{
  "schema_version": "vision_question_clusters_resolved_v1",
  "dataset_family": "image_classification",
  "selection_metric": "test_cross_entropy",
  "profile_prefix": "image_classification_template_",
  "source_path": "profiles/image-classification-ranking-clusters-v1.yaml",
  "source_sha256": "a9f978cb00d7650c29c7389e5fce4b6945945e727c66ce2ca22de6e2b9a2c627",
  "base_profile": "profiles/task-expansion-image-classification.yaml",
  "default_num_choices": 3,
  "discovery_training_seeds": [
    20,
    21,
    22,
    23
  ],
  "confirmation_training_seeds": [
    200,
    201,
    202,
    203,
    204,
    205,
    206,
    207,
    208,
    209,
    210,
    211,
    212,
    213,
    214,
    215
  ],
  "selection_contract": {
    "discovery_split": "validation",
    "confirmation_split": "test",
    "freeze_before_confirmation": [
      "candidate IDs",
      "adjacent ranking directions",
      "comparison family",
      "gap threshold",
      "seed list"
    ],
    "multiplicity": "Holm across every frozen adjacent hypothesis, including rejected templates",
    "minimum_mean_gap": 0.005,
    "require_nonoverlap": false,
    "claim_scope": "Synthetic task and fixed image-instance conditional training seed robustness. No natural-image claim.",
    "difficulty_policy": "Evaluate every admitted triple; retain rejected triples without replacement."
  },
  "statistics": {
    "discovery": {
      "n_seeds": 4,
      "split": "validation",
      "purpose": "direction_estimation_only",
      "ranking_rule": "ascending_mean_validation_cross_entropy",
      "significance_gate": false
    },
    "confirmation": {
      "n_seeds": 16,
      "split": "test",
      "test": "one_sided_paired_sign_test",
      "alpha": 0.05,
      "correction": "holm",
      "min_mean_paired_gap": 0.005
    },
    "estimand": "Paired per-seed win/loss stability plus observed mean CE gap for a fixed materialized dataset. Not a test of expected CE differences or natural-image generalization."
  },
  "protocol_config": {
    "schema_version": "vision_ranking_protocol_v1",
    "discovery": {
      "seeds": [
        20,
        21,
        22,
        23
      ],
      "metric": "final_validation_metric",
      "order": "ascending mean; ties use candidate_id",
      "purpose": "Freeze a direction for every template instance before independent confirmation. Weak discovery separation is retained and reported, not an automatic rejection."
    },
    "confirmation": {
      "seeds": [
        200,
        201,
        202,
        203,
        204,
        205,
        206,
        207,
        208,
        209,
        210,
        211,
        212,
        213,
        214,
        215
      ],
      "metric": "final_test_cross_entropy",
      "higher_is_better": false,
      "test": "one-sided exact paired binomial sign test; ties count as nonpositive",
      "null": "P(oriented paired improvement > 0) <= 0.5 over training seed pairs, conditional on fixed materialized data and training protocol",
      "multiplicity": "Holm across every frozen adjacent comparison, including failed or missing comparisons with p=1",
      "alpha": 0.05,
      "minimum_adjacent_mean_gap": 0.005,
      "require_nonoverlapping_seed_ranges": false,
      "bootstrap": {
        "samples": 10000,
        "confidence": 0.95,
        "seed": 73013,
        "purpose": "Descriptive paired mean-gap interval only; no admission decision uses this interval"
      },
      "failure_policy": "Any absent, duplicate, nonfinite or failed seed rejects that instance. No seed deletion, replacement, added seeds, reordered choices, or extra metric selection after confirmation."
    },
    "admission": "Every adjacent comparison in the discovery-frozen order must pass Holm and the minimum mean-gap gate; both gates use oriented improvement in the frozen direction.",
    "independence": "Confirmation seeds must be disjoint from discovery seeds. Freeze full membership, candidate order, artifact hashes, metrics, gates, and hypothesis family before starting confirmation. Never select questions using tested models responses.",
    "comparison_contract": {
      "same_data": "Every choice shares one dataset instance and identical split file hashes.",
      "budgets": "Equal total_samples_seen unless the cluster explicitly intervenes on budget; always expose steps, batch size and total_samples_seen per choice. Equal samples with different batches does not imply equal optimizer updates.",
      "parameters": "Record measured parameter count. Parameter matching is a separate intervention, never inferred from width matching.",
      "streams": "Use paired training seeds and shared batch index stream when batch sizes agree; seed number alone does not establish matched initialization across architectures.",
      "interventions": "Record every changed spec field and final full-code hashes. Single-axis interpretation requires other fields fixed; multi-axis choices establish joint effects only.",
      "devices": [
        "360-1",
        "ophis"
      ]
    },
    "evaluation": {
      "models": [
        "claude-opus-5-5",
        "gpt-6.1-sol",
        "claude-haiku-5-5",
        "gpt-6-luna"
      ],
      "reasoning_effort": "medium",
      "max_tokens": 8192,
      "tools": false,
      "calls_per_question": 1,
      "system": "Answer the supplied question. Tools are unavailable.",
      "response": "Public explanation followed by exact full ordering. Audit explanation independently of correctness.",
      "metrics": [
        "exact full order",
        "pairwise concordance",
        "valid response rate",
        "answer-explanation consistency"
      ],
      "failure_policy": "Only transport failures without completed answer may retry. Preserve all attempts, identity, prompt hashes and public text. Never replace a completed wrong answer.",
      "baseline": "For k choices, uniform complete-ranking probability is 1/k!; pairwise chance is 1/2.",
      "prompt": "Use canonical renderer with complete shared base and independent diffs; byte-exact reconstruction checked. Fallback to full code when diff readability degrades."
    },
    "difficulty": {
      "surface_heuristics": "Freeze interpretable surface-only rules before LLM responses; report abstentions and tie coverage. Do not use confirmation CE or model errors to tune rules.",
      "coverage": "Report every frozen template instance as accepted/rejected/failed and why. Do not only publish difficult successes.",
      "dependence": "Report results by template and teacher draw. Repeated teacher or dataset conditions are correlated; no independent-question binomial intervals or general model ranking claims.",
      "selection": "Do not use model mistakes to admit or drop items. Any later difficult subset is explicitly exploratory and kept separate from full predeclared batch."
    },
    "causal_scope": "Confirmed performance differences identify the total effect of the exposed intervention at fixed data and training horizon. A mechanism is supported only to the extent that separately logged diagnostic/control evidence rules out alternatives. Image results concern the stated synthetic image distribution only.",
    "statistics": {
      "discovery": {
        "n_seeds": 4,
        "split": "validation",
        "purpose": "direction_estimation_only",
        "ranking_rule": "ascending_mean_validation_cross_entropy",
        "significance_gate": false
      },
      "confirmation": {
        "n_seeds": 16,
        "split": "test",
        "test": "one_sided_paired_sign_test",
        "alpha": 0.05,
        "correction": "holm",
        "min_mean_paired_gap": 0.005
      },
      "estimand": "Paired per-seed win/loss stability plus observed mean CE gap for a fixed materialized dataset. Not a test of expected CE differences or natural-image generalization."
    },
    "answer_format": "best_to_worst_v1"
  },
  "surface_heuristics": [],
  "prompt_contract": {
    "representation": "shared_complete_base_plus_independent_diff",
    "answer_format": "best_to_worst_v1",
    "analyst_only_fields": [
      "hypothesis",
      "shortcut_checks",
      "explanation_target",
      "admission"
    ]
  },
  "worker_sources": [
    "tools/ntp_h200_worker.py",
    "tools/ntp_anchor_worker.py",
    "tools/vision_h200_worker.py"
  ],
  "counts": {
    "templates": 18,
    "instances": 54,
    "dataset_nodes": 15,
    "candidate_nodes": 207,
    "distinct_models": 42
  },
  "dataset_nodes": [
    {
      "node_id": "dataset_e0bb4b3a905a5977",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841901,
      "context_ref": "relations"
    },
    {
      "node_id": "dataset_f7fa567d4761b40c",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841902,
      "context_ref": "relations"
    },
    {
      "node_id": "dataset_1544b3cbd88f855f",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841903,
      "context_ref": "relations"
    },
    {
      "node_id": "dataset_68a832a939c0b770",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841901,
      "context_ref": "shapes"
    },
    {
      "node_id": "dataset_acbb3cacc35c6e86",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841902,
      "context_ref": "shapes"
    },
    {
      "node_id": "dataset_43d6af5f6d3b729d",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841903,
      "context_ref": "shapes"
    },
    {
      "node_id": "dataset_9f616ad5ea615654",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841901,
      "context_ref": "textures"
    },
    {
      "node_id": "dataset_9379541427e4f9a7",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841902,
      "context_ref": "textures"
    },
    {
      "node_id": "dataset_95967aad9ad199bb",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841903,
      "context_ref": "textures"
    },
    {
      "node_id": "dataset_32df760ef312a38d",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          4.0,
          6.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841901,
      "context_ref": "fine_textures"
    },
    {
      "node_id": "dataset_e6c4261755f0ff04",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          4.0,
          6.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841902,
      "context_ref": "fine_textures"
    },
    {
      "node_id": "dataset_220eb81b568bc599",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          4.0,
          6.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841903,
      "context_ref": "fine_textures"
    },
    {
      "node_id": "dataset_4d28614559aee4b9",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 64,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841901,
      "context_ref": "small_shapes"
    },
    {
      "node_id": "dataset_484bdc518daa14bb",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 64,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841902,
      "context_ref": "small_shapes"
    },
    {
      "node_id": "dataset_63f24d2a7f7dd6e2",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 64,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "instance_seed": 841903,
      "context_ref": "small_shapes"
    }
  ],
  "candidate_nodes": [
    {
      "node_id": "candidate_aedc559262a997543320",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 1,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 8572,
      "model_code_sha256": "2744aac8d669a5a4171096c59dc5d68686ab43164fc80bec2b0f91f9c576457d"
    },
    {
      "node_id": "candidate_e2e98633894e1a273735",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 4,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 31468,
      "model_code_sha256": "d6ce0d26e9f8f7790ae231b01bdd44e8e90cd75d74db93df02800810ecc27204"
    },
    {
      "node_id": "candidate_bb23efca29cbc84e0f54",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 1,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 31612,
      "model_code_sha256": "ddc392f9bdd8df280f3822a219a6653dbbef09d29aa051a52ca5fae1ec2ed0e1"
    },
    {
      "node_id": "candidate_96057a489f621815967a",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 4,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 123628,
      "model_code_sha256": "de12e3d1b8c3c5d6f6b890a1b9e1449f76bb899fd1b6db4d44d794f99ab91150"
    },
    {
      "node_id": "candidate_5c73c2b705cd3ac05e84",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 1,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 8572,
      "model_code_sha256": "2744aac8d669a5a4171096c59dc5d68686ab43164fc80bec2b0f91f9c576457d"
    },
    {
      "node_id": "candidate_82928db599c08a734f6d",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 4,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 31468,
      "model_code_sha256": "d6ce0d26e9f8f7790ae231b01bdd44e8e90cd75d74db93df02800810ecc27204"
    },
    {
      "node_id": "candidate_264ad6f33a2d476f3ffd",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 1,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 31612,
      "model_code_sha256": "ddc392f9bdd8df280f3822a219a6653dbbef09d29aa051a52ca5fae1ec2ed0e1"
    },
    {
      "node_id": "candidate_c241ce5e2ffc71265de4",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 4,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 123628,
      "model_code_sha256": "de12e3d1b8c3c5d6f6b890a1b9e1449f76bb899fd1b6db4d44d794f99ab91150"
    },
    {
      "node_id": "candidate_016ee74c9be9565d2e2f",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 1,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 8572,
      "model_code_sha256": "2744aac8d669a5a4171096c59dc5d68686ab43164fc80bec2b0f91f9c576457d"
    },
    {
      "node_id": "candidate_0e6bb18a73baeeb2cd6e",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 4,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 31468,
      "model_code_sha256": "d6ce0d26e9f8f7790ae231b01bdd44e8e90cd75d74db93df02800810ecc27204"
    },
    {
      "node_id": "candidate_be9ddfc1b3d86605ec66",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 1,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 31612,
      "model_code_sha256": "ddc392f9bdd8df280f3822a219a6653dbbef09d29aa051a52ca5fae1ec2ed0e1"
    },
    {
      "node_id": "candidate_d8dfd33851eb7eb58bf5",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 4,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 123628,
      "model_code_sha256": "de12e3d1b8c3c5d6f6b890a1b9e1449f76bb899fd1b6db4d44d794f99ab91150"
    },
    {
      "node_id": "candidate_3cbaed19cf1f1bd78892",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_fe248f3711cecdf66c13",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": true,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15484,
      "model_code_sha256": "b7f80d8dbdfffa7aca4b82e112df579d6be15d78c5668c7e516661f5532af7d1"
    },
    {
      "node_id": "candidate_675b0f0945a65c7728d6",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_4e616725a8ecc13c433b",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": true,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16252,
      "model_code_sha256": "f105901133fe76424f8e503f9cc6cca6a84d8017371e00f46f2fde2d43c11841"
    },
    {
      "node_id": "candidate_c93709ee26aacbc9f437",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_d91d654be99bc328c040",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": true,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15484,
      "model_code_sha256": "b7f80d8dbdfffa7aca4b82e112df579d6be15d78c5668c7e516661f5532af7d1"
    },
    {
      "node_id": "candidate_f9bdfa11ef52bc5e3528",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_a644642ee0d58a7cbf09",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": true,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16252,
      "model_code_sha256": "f105901133fe76424f8e503f9cc6cca6a84d8017371e00f46f2fde2d43c11841"
    },
    {
      "node_id": "candidate_07cf9b9e2ef0417e1afc",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_c65bf21470c8b1e81842",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": true,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15484,
      "model_code_sha256": "b7f80d8dbdfffa7aca4b82e112df579d6be15d78c5668c7e516661f5532af7d1"
    },
    {
      "node_id": "candidate_d6f536a09652df88c280",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_4b043b424317b46e875a",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": true,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16252,
      "model_code_sha256": "f105901133fe76424f8e503f9cc6cca6a84d8017371e00f46f2fde2d43c11841"
    },
    {
      "node_id": "candidate_cb13dd586f5542e7de21",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 3,
        "position": "none",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 17828,
      "model_code_sha256": "9defa266b5335655ef99314c9606112a78f37e8f7ebf5c14b80373e59e30b747"
    },
    {
      "node_id": "candidate_5cf50db65aa67995a4f9",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_f40fb6fc801f6d2a3c95",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 6,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20068,
      "model_code_sha256": "e71b2de720715f358b0a505a393993f429c24e76a254225b74c00743a9c38227"
    },
    {
      "node_id": "candidate_e7f426dc4d829a438fe9",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 3,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19876,
      "model_code_sha256": "f29b6e6a38c3792bb771f0a25777e143f7555c8c6cf80cd1cd0007277960f5b3"
    },
    {
      "node_id": "candidate_159ce277330e0990eb83",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "none",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 18276,
      "model_code_sha256": "fbe5f23efc3b8d68938d38f8539601ea9f2b82f6121f23d6f218d9256e9bb082"
    },
    {
      "node_id": "candidate_970c26429b51319c117b",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 6,
        "position": "none",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19556,
      "model_code_sha256": "13e3d8cca256e65b2821e79f624f79f2ea20c5314d05efe3f9e68aa9b4913a35"
    },
    {
      "node_id": "candidate_9096f1890ddea0000c15",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 3,
        "position": "none",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 17828,
      "model_code_sha256": "9defa266b5335655ef99314c9606112a78f37e8f7ebf5c14b80373e59e30b747"
    },
    {
      "node_id": "candidate_5313f05220195b0f67e9",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_76dacecff1eed9e2c8e8",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 6,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20068,
      "model_code_sha256": "e71b2de720715f358b0a505a393993f429c24e76a254225b74c00743a9c38227"
    },
    {
      "node_id": "candidate_7cb8c5ca7cbb9b8fd732",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 3,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19876,
      "model_code_sha256": "f29b6e6a38c3792bb771f0a25777e143f7555c8c6cf80cd1cd0007277960f5b3"
    },
    {
      "node_id": "candidate_5314c84c867a5b81a5ea",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "none",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 18276,
      "model_code_sha256": "fbe5f23efc3b8d68938d38f8539601ea9f2b82f6121f23d6f218d9256e9bb082"
    },
    {
      "node_id": "candidate_41c996694bfdeb01d4eb",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 6,
        "position": "none",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19556,
      "model_code_sha256": "13e3d8cca256e65b2821e79f624f79f2ea20c5314d05efe3f9e68aa9b4913a35"
    },
    {
      "node_id": "candidate_365c90b1233692d34e21",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 3,
        "position": "none",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 17828,
      "model_code_sha256": "9defa266b5335655ef99314c9606112a78f37e8f7ebf5c14b80373e59e30b747"
    },
    {
      "node_id": "candidate_681cbfb92210830d5199",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_dfb5edbd1f2cbf291954",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 6,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20068,
      "model_code_sha256": "e71b2de720715f358b0a505a393993f429c24e76a254225b74c00743a9c38227"
    },
    {
      "node_id": "candidate_98e22078313daa99d5d2",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 3,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19876,
      "model_code_sha256": "f29b6e6a38c3792bb771f0a25777e143f7555c8c6cf80cd1cd0007277960f5b3"
    },
    {
      "node_id": "candidate_8c676e89972ddefe6f92",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "none",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 18276,
      "model_code_sha256": "fbe5f23efc3b8d68938d38f8539601ea9f2b82f6121f23d6f218d9256e9bb082"
    },
    {
      "node_id": "candidate_4866c13502bb9e784e3e",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 6,
        "position": "none",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19556,
      "model_code_sha256": "13e3d8cca256e65b2821e79f624f79f2ea20c5314d05efe3f9e68aa9b4913a35"
    },
    {
      "node_id": "candidate_9fe585c7b9d737d48244",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "mlp",
        "width": 96,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 129700,
      "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010"
    },
    {
      "node_id": "candidate_eb0fc9a5ce2767f19d93",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "mlp",
        "width": 96,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 129700,
      "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010"
    },
    {
      "node_id": "candidate_32c63bc42039bb009be4",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "mlp",
        "width": 96,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 129700,
      "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010"
    },
    {
      "node_id": "candidate_3b74befeb2c6c9eb61c9",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_5387a55740ddb1ad3894",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "mlp",
        "width": 96,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 129700,
      "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010"
    },
    {
      "node_id": "candidate_ac04e963dc5200162093",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_c5a024c8f4d8af48cf51",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_7399cba3f8ca90f400cc",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_243be9f0f4d2817e59a6",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "mlp",
        "width": 96,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 129700,
      "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010"
    },
    {
      "node_id": "candidate_6cb11cedc2472110e7b0",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_c79b5e3e8168a286054c",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_ec941ffe7c1629588866",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_0042c2c253ff4d460405",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "mlp",
        "width": 96,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 129700,
      "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010"
    },
    {
      "node_id": "candidate_b8f7de70a865c14a6fc0",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_bfd304b5b5806dae106f",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_ab616a897d11be58391b",
      "dataset_node_id": "dataset_9f616ad5ea615654",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_149005313738d78a4ed7",
      "dataset_node_id": "dataset_9f616ad5ea615654",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 5,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 2,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 34012,
      "model_code_sha256": "394a7050f24d49bb1c1fe5a84fddde7fb349d74bb928ebd3d56918f08922ce9e"
    },
    {
      "node_id": "candidate_aafda8e15a0bbf01406c",
      "dataset_node_id": "dataset_9f616ad5ea615654",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 4,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16156,
      "model_code_sha256": "247e1e707578507112126325851aaa80fe6e7e885427ce689c887595d7b6e215"
    },
    {
      "node_id": "candidate_6e5f0ab5730336704ac8",
      "dataset_node_id": "dataset_9f616ad5ea615654",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 5,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 33868,
      "model_code_sha256": "e9a9a2765f20d57dccca28a3fffdc7dadefe39637b32ed6013d958be3159f651"
    },
    {
      "node_id": "candidate_a434611033b8affdc6bd",
      "dataset_node_id": "dataset_9f616ad5ea615654",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 2,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15580,
      "model_code_sha256": "4eed0b51c7ee7ec1f74effe0bdf240b1ba24009b9af627dd4c278268ccdac078"
    },
    {
      "node_id": "candidate_4199c0cb23d1c4c4703a",
      "dataset_node_id": "dataset_9f616ad5ea615654",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 5,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 4,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 34588,
      "model_code_sha256": "df29ae51767411dcd43886969870251fe836bfde85d67f093e5d3a629a91112d"
    },
    {
      "node_id": "candidate_fad13efd0f5a1af0c4eb",
      "dataset_node_id": "dataset_9f616ad5ea615654",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_37f73a292361df1c6720",
      "dataset_node_id": "dataset_9379541427e4f9a7",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_66d46963eeb75fd48934",
      "dataset_node_id": "dataset_9379541427e4f9a7",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 5,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 2,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 34012,
      "model_code_sha256": "394a7050f24d49bb1c1fe5a84fddde7fb349d74bb928ebd3d56918f08922ce9e"
    },
    {
      "node_id": "candidate_f430ce2ad178a479ae16",
      "dataset_node_id": "dataset_9379541427e4f9a7",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 4,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16156,
      "model_code_sha256": "247e1e707578507112126325851aaa80fe6e7e885427ce689c887595d7b6e215"
    },
    {
      "node_id": "candidate_abcdf25ddb5ee63b8869",
      "dataset_node_id": "dataset_9379541427e4f9a7",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 5,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 33868,
      "model_code_sha256": "e9a9a2765f20d57dccca28a3fffdc7dadefe39637b32ed6013d958be3159f651"
    },
    {
      "node_id": "candidate_c0578c1bb16eb1b7a0df",
      "dataset_node_id": "dataset_9379541427e4f9a7",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 2,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15580,
      "model_code_sha256": "4eed0b51c7ee7ec1f74effe0bdf240b1ba24009b9af627dd4c278268ccdac078"
    },
    {
      "node_id": "candidate_4934024985276952b5ab",
      "dataset_node_id": "dataset_9379541427e4f9a7",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 5,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 4,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 34588,
      "model_code_sha256": "df29ae51767411dcd43886969870251fe836bfde85d67f093e5d3a629a91112d"
    },
    {
      "node_id": "candidate_6ed7409ffc8cf32c983f",
      "dataset_node_id": "dataset_9379541427e4f9a7",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_db0952c46eff7da757dd",
      "dataset_node_id": "dataset_95967aad9ad199bb",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_7330b39f40c0868f3f45",
      "dataset_node_id": "dataset_95967aad9ad199bb",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 5,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 2,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 34012,
      "model_code_sha256": "394a7050f24d49bb1c1fe5a84fddde7fb349d74bb928ebd3d56918f08922ce9e"
    },
    {
      "node_id": "candidate_b017aaec9762ac9f8f15",
      "dataset_node_id": "dataset_95967aad9ad199bb",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 4,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16156,
      "model_code_sha256": "247e1e707578507112126325851aaa80fe6e7e885427ce689c887595d7b6e215"
    },
    {
      "node_id": "candidate_8e91b1855c6598f64aa6",
      "dataset_node_id": "dataset_95967aad9ad199bb",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 5,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 33868,
      "model_code_sha256": "e9a9a2765f20d57dccca28a3fffdc7dadefe39637b32ed6013d958be3159f651"
    },
    {
      "node_id": "candidate_5d1deba868feca1122eb",
      "dataset_node_id": "dataset_95967aad9ad199bb",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 2,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15580,
      "model_code_sha256": "4eed0b51c7ee7ec1f74effe0bdf240b1ba24009b9af627dd4c278268ccdac078"
    },
    {
      "node_id": "candidate_c26909b9e47e00ad719f",
      "dataset_node_id": "dataset_95967aad9ad199bb",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 5,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 4,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 34588,
      "model_code_sha256": "df29ae51767411dcd43886969870251fe836bfde85d67f093e5d3a629a91112d"
    },
    {
      "node_id": "candidate_0bc372c1935b61980671",
      "dataset_node_id": "dataset_95967aad9ad199bb",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_19540c3e091616ae57fb",
      "dataset_node_id": "dataset_32df760ef312a38d",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 3,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19876,
      "model_code_sha256": "f29b6e6a38c3792bb771f0a25777e143f7555c8c6cf80cd1cd0007277960f5b3"
    },
    {
      "node_id": "candidate_3832b1a5bed99bc11e88",
      "dataset_node_id": "dataset_32df760ef312a38d",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_c36f86452a6409a9d0f0",
      "dataset_node_id": "dataset_32df760ef312a38d",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 6,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20068,
      "model_code_sha256": "e71b2de720715f358b0a505a393993f429c24e76a254225b74c00743a9c38227"
    },
    {
      "node_id": "candidate_d388f61a539fdadaee0a",
      "dataset_node_id": "dataset_e6c4261755f0ff04",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 3,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19876,
      "model_code_sha256": "f29b6e6a38c3792bb771f0a25777e143f7555c8c6cf80cd1cd0007277960f5b3"
    },
    {
      "node_id": "candidate_c65741c175d9e7f58b17",
      "dataset_node_id": "dataset_e6c4261755f0ff04",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_7b66ca9bca5de7458b8f",
      "dataset_node_id": "dataset_e6c4261755f0ff04",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 6,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20068,
      "model_code_sha256": "e71b2de720715f358b0a505a393993f429c24e76a254225b74c00743a9c38227"
    },
    {
      "node_id": "candidate_ee1139b5d486a75aeb5b",
      "dataset_node_id": "dataset_220eb81b568bc599",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 3,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19876,
      "model_code_sha256": "f29b6e6a38c3792bb771f0a25777e143f7555c8c6cf80cd1cd0007277960f5b3"
    },
    {
      "node_id": "candidate_3afa0c1c9fcc541e8993",
      "dataset_node_id": "dataset_220eb81b568bc599",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_a0b806c0a1e8a7e9d67a",
      "dataset_node_id": "dataset_220eb81b568bc599",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 4,
        "patch_size": 6,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20068,
      "model_code_sha256": "e71b2de720715f358b0a505a393993f429c24e76a254225b74c00743a9c38227"
    },
    {
      "node_id": "candidate_451e3455c06ccd985fa1",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "depthwise",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 5500,
      "model_code_sha256": "297bdbdefb2459efc98a425849209a52ad5c0084bb3bc416d73fb096109e7e4e"
    },
    {
      "node_id": "candidate_bcb7d96501c45ac8fc9b",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 48,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "depthwise",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20212,
      "model_code_sha256": "c5b04c3c387f9d97858c89cc93dd742f98b4fc635d9ddef274224866d4a35f0a"
    },
    {
      "node_id": "candidate_ccb91e6df29450d94d92",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 48,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 60820,
      "model_code_sha256": "751ce9f6d2824bfbe26fced8a343596c9f2f627a889f39ed9cd667537564058c"
    },
    {
      "node_id": "candidate_bce264253457a26bed58",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "depthwise",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 5500,
      "model_code_sha256": "297bdbdefb2459efc98a425849209a52ad5c0084bb3bc416d73fb096109e7e4e"
    },
    {
      "node_id": "candidate_47be65ef0493e655d027",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 48,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "depthwise",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20212,
      "model_code_sha256": "c5b04c3c387f9d97858c89cc93dd742f98b4fc635d9ddef274224866d4a35f0a"
    },
    {
      "node_id": "candidate_090457311c5d735692bb",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 48,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 60820,
      "model_code_sha256": "751ce9f6d2824bfbe26fced8a343596c9f2f627a889f39ed9cd667537564058c"
    },
    {
      "node_id": "candidate_a1697d020b7aceb34212",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "depthwise",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 5500,
      "model_code_sha256": "297bdbdefb2459efc98a425849209a52ad5c0084bb3bc416d73fb096109e7e4e"
    },
    {
      "node_id": "candidate_54f7bf4b45b4f2cc5dba",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 48,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "depthwise",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20212,
      "model_code_sha256": "c5b04c3c387f9d97858c89cc93dd742f98b4fc635d9ddef274224866d4a35f0a"
    },
    {
      "node_id": "candidate_ebe867dbb00d2d455551",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 48,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 60820,
      "model_code_sha256": "751ce9f6d2824bfbe26fced8a343596c9f2f627a889f39ed9cd667537564058c"
    },
    {
      "node_id": "candidate_5e02b72d5e785e09d855",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "487f99a81970de56cd83e5ed249732dbfa7860e0a5389abfec6823e543153207"
    },
    {
      "node_id": "candidate_d970952e0279a86f3e85",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 6,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 46732,
      "model_code_sha256": "7183d61f2bf2188b283e14ce82318d20f5855fcd0a16bc4a57456c2081c89b67"
    },
    {
      "node_id": "candidate_ba0e44517b74430fa307",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 6,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 46732,
      "model_code_sha256": "7818c319584d6ddd071c6fa738ca50c22811f9f460d2e6e4469a54be52606813"
    },
    {
      "node_id": "candidate_c9945d9e94130fa6d514",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "487f99a81970de56cd83e5ed249732dbfa7860e0a5389abfec6823e543153207"
    },
    {
      "node_id": "candidate_9dd05632feeef8396a70",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 6,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 46732,
      "model_code_sha256": "7183d61f2bf2188b283e14ce82318d20f5855fcd0a16bc4a57456c2081c89b67"
    },
    {
      "node_id": "candidate_3d75613e820affe88db9",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 6,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 46732,
      "model_code_sha256": "7818c319584d6ddd071c6fa738ca50c22811f9f460d2e6e4469a54be52606813"
    },
    {
      "node_id": "candidate_ec18f033d414ed8b2267",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "487f99a81970de56cd83e5ed249732dbfa7860e0a5389abfec6823e543153207"
    },
    {
      "node_id": "candidate_6df66b35fc95baca2b92",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 6,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 46732,
      "model_code_sha256": "7183d61f2bf2188b283e14ce82318d20f5855fcd0a16bc4a57456c2081c89b67"
    },
    {
      "node_id": "candidate_ce554a6b3b83b800b884",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 6,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 46732,
      "model_code_sha256": "7818c319584d6ddd071c6fa738ca50c22811f9f460d2e6e4469a54be52606813"
    },
    {
      "node_id": "candidate_bb0449ecad9dc1b5fc11",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "none",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15340,
      "model_code_sha256": "384ea3383ad48c29b344417d8d4ed4a0692433ae1e3d913dceeb3f2764b1ed74"
    },
    {
      "node_id": "candidate_1e525757e242e2130e66",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "c08a63248209b0d5a64c9236d54c5be7f1beb598dc26e8c73271614ded80629b"
    },
    {
      "node_id": "candidate_6b94f7549d0848a00d04",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": false,
        "norm": "none",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15340,
      "model_code_sha256": "ea5fa570ff31093c95f665efc19d7c1108483d3113df92e6417e1a5f1fe6a40e"
    },
    {
      "node_id": "candidate_847a0405ced8043e5e4f",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "none",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15340,
      "model_code_sha256": "384ea3383ad48c29b344417d8d4ed4a0692433ae1e3d913dceeb3f2764b1ed74"
    },
    {
      "node_id": "candidate_987bbf0bc6a81712d9f8",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "c08a63248209b0d5a64c9236d54c5be7f1beb598dc26e8c73271614ded80629b"
    },
    {
      "node_id": "candidate_9a9a9b9d6908cdc857fc",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": false,
        "norm": "none",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15340,
      "model_code_sha256": "ea5fa570ff31093c95f665efc19d7c1108483d3113df92e6417e1a5f1fe6a40e"
    },
    {
      "node_id": "candidate_fa16bd42cc361c317d1e",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "none",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15340,
      "model_code_sha256": "384ea3383ad48c29b344417d8d4ed4a0692433ae1e3d913dceeb3f2764b1ed74"
    },
    {
      "node_id": "candidate_979cc6b964dc1ad05d7a",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "c08a63248209b0d5a64c9236d54c5be7f1beb598dc26e8c73271614ded80629b"
    },
    {
      "node_id": "candidate_dc7213f3bdc686d92450",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": false,
        "norm": "none",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15340,
      "model_code_sha256": "ea5fa570ff31093c95f665efc19d7c1108483d3113df92e6417e1a5f1fe6a40e"
    },
    {
      "node_id": "candidate_b761d5a572e7c8930b00",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "relu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 1,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 13084,
      "model_code_sha256": "d9997a6866de1641a44db88c5d3ea82cbc449f26a4ffc4d05ac1a573bbcb0a30"
    },
    {
      "node_id": "candidate_99cb23fdd2ff81139333",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "relu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 4,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20140,
      "model_code_sha256": "c3cac7123a9b71d029392ad0bd01da692d128efe5f3f2267300a9d4cd61e5659"
    },
    {
      "node_id": "candidate_121eb45f3650881324a7",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 1,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 13084,
      "model_code_sha256": "1486f09e036eb722ea94783ce64d303a75e37f0d35a5b0010bf6d3362b2bb66c"
    },
    {
      "node_id": "candidate_64fb301b46a587e3fc10",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "relu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "a93226785f393655cecb8174ae617a22d94e8251a2ae6e3ea53ad51c9e5b4e75"
    },
    {
      "node_id": "candidate_a4320ec04055e927c458",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 4,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20140,
      "model_code_sha256": "32cca09dafc2c833b1906f5574e0c99166b7f54576f13dae140418f49f45a2cb"
    },
    {
      "node_id": "candidate_6b81d3cd60d9b34c45c9",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "relu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 1,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 13084,
      "model_code_sha256": "d9997a6866de1641a44db88c5d3ea82cbc449f26a4ffc4d05ac1a573bbcb0a30"
    },
    {
      "node_id": "candidate_3692b717bd7aa351ee60",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "relu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 4,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20140,
      "model_code_sha256": "c3cac7123a9b71d029392ad0bd01da692d128efe5f3f2267300a9d4cd61e5659"
    },
    {
      "node_id": "candidate_69446f8cd2138a6a629e",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 1,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 13084,
      "model_code_sha256": "1486f09e036eb722ea94783ce64d303a75e37f0d35a5b0010bf6d3362b2bb66c"
    },
    {
      "node_id": "candidate_614911c0fd261ee640a2",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "relu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "a93226785f393655cecb8174ae617a22d94e8251a2ae6e3ea53ad51c9e5b4e75"
    },
    {
      "node_id": "candidate_f9a29ada749d4adf5672",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 4,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20140,
      "model_code_sha256": "32cca09dafc2c833b1906f5574e0c99166b7f54576f13dae140418f49f45a2cb"
    },
    {
      "node_id": "candidate_516e799a1c22179b8da9",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "relu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 1,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 13084,
      "model_code_sha256": "d9997a6866de1641a44db88c5d3ea82cbc449f26a4ffc4d05ac1a573bbcb0a30"
    },
    {
      "node_id": "candidate_a51f41b8733efb67e1dd",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "relu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 4,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20140,
      "model_code_sha256": "c3cac7123a9b71d029392ad0bd01da692d128efe5f3f2267300a9d4cd61e5659"
    },
    {
      "node_id": "candidate_102482d9d6b525ece005",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 1,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 13084,
      "model_code_sha256": "1486f09e036eb722ea94783ce64d303a75e37f0d35a5b0010bf6d3362b2bb66c"
    },
    {
      "node_id": "candidate_f55c23b4528fe92a3654",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "relu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "a93226785f393655cecb8174ae617a22d94e8251a2ae6e3ea53ad51c9e5b4e75"
    },
    {
      "node_id": "candidate_8c794a0d7e554257b434",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 4,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 20140,
      "model_code_sha256": "32cca09dafc2c833b1906f5574e0c99166b7f54576f13dae140418f49f45a2cb"
    },
    {
      "node_id": "candidate_d315cbf6c5d694184220",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 4,
        "heads": 1,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 27748,
      "model_code_sha256": "33aa4418bafc50c1f23adb582ed959f04f2cfe8daf97db6cd775f669564215b7"
    },
    {
      "node_id": "candidate_d0e587aed7b3c0b52997",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 1,
        "heads": 8,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15268,
      "model_code_sha256": "37b0b123040119e40aa10bdfce698d951381255500bad19d4d734cb820a3bc1d"
    },
    {
      "node_id": "candidate_f1ea60b4a966c158b63c",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 1,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "4038255caa48fa6b4fa7829427ab55fcf105ad98de0ac407c33ee34def5a043a"
    },
    {
      "node_id": "candidate_8630080f46d5bcc4c4fa",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 8,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "9e76b05e6a7a0f7d75de4c9fd9ac4652a7ce37e4f12081897317f660e7ca7621"
    },
    {
      "node_id": "candidate_699c6c5c2d2eb10b3040",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 1,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15268,
      "model_code_sha256": "8364bbae844edba18d6067aa23e9bf7faaded3ed3722211383bbd3e63c55c6da"
    },
    {
      "node_id": "candidate_cdfd16b183904fe63c5a",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 4,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 27748,
      "model_code_sha256": "5e4a2c32ed8426dea1fabecda2b813b8ff8e4a7d0fcafbd8be819520e625e953"
    },
    {
      "node_id": "candidate_2e250ec49da02dabf3e9",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 4,
        "heads": 1,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 27748,
      "model_code_sha256": "33aa4418bafc50c1f23adb582ed959f04f2cfe8daf97db6cd775f669564215b7"
    },
    {
      "node_id": "candidate_09a8640fa4d7c593644f",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 1,
        "heads": 8,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15268,
      "model_code_sha256": "37b0b123040119e40aa10bdfce698d951381255500bad19d4d734cb820a3bc1d"
    },
    {
      "node_id": "candidate_2d28083738832680f9be",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 1,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "4038255caa48fa6b4fa7829427ab55fcf105ad98de0ac407c33ee34def5a043a"
    },
    {
      "node_id": "candidate_2e24746f420e07884c4c",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 8,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "9e76b05e6a7a0f7d75de4c9fd9ac4652a7ce37e4f12081897317f660e7ca7621"
    },
    {
      "node_id": "candidate_a301024fb64cd6869bd2",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 1,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15268,
      "model_code_sha256": "8364bbae844edba18d6067aa23e9bf7faaded3ed3722211383bbd3e63c55c6da"
    },
    {
      "node_id": "candidate_db3bac963e47c6104d6a",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 4,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 27748,
      "model_code_sha256": "5e4a2c32ed8426dea1fabecda2b813b8ff8e4a7d0fcafbd8be819520e625e953"
    },
    {
      "node_id": "candidate_c205e04736f94326026a",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 4,
        "heads": 1,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 27748,
      "model_code_sha256": "33aa4418bafc50c1f23adb582ed959f04f2cfe8daf97db6cd775f669564215b7"
    },
    {
      "node_id": "candidate_b42338d48a1242eb97d9",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 1,
        "heads": 8,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15268,
      "model_code_sha256": "37b0b123040119e40aa10bdfce698d951381255500bad19d4d734cb820a3bc1d"
    },
    {
      "node_id": "candidate_58d373fe55636c9acd59",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 1,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "4038255caa48fa6b4fa7829427ab55fcf105ad98de0ac407c33ee34def5a043a"
    },
    {
      "node_id": "candidate_229ab15aa9cbfc005fb3",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "heads": 8,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "9e76b05e6a7a0f7d75de4c9fd9ac4652a7ce37e4f12081897317f660e7ca7621"
    },
    {
      "node_id": "candidate_c0eff0c920f37eab36d3",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 1,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15268,
      "model_code_sha256": "8364bbae844edba18d6067aa23e9bf7faaded3ed3722211383bbd3e63c55c6da"
    },
    {
      "node_id": "candidate_515b18e0245f3d9b63bd",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "patch_transformer",
        "width": 32,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 4,
        "heads": 4,
        "patch_size": 4,
        "position": "learned",
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 300,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          75,
          150,
          300
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 27748,
      "model_code_sha256": "5e4a2c32ed8426dea1fabecda2b813b8ff8e4a7d0fcafbd8be819520e625e953"
    },
    {
      "node_id": "candidate_31abb7f5ae858e9f35e8",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.0001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_6ec59d4186f9b42d6754",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_b07779969f06e2a919a4",
      "dataset_node_id": "dataset_68a832a939c0b770",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.01,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_68c3418242c7dd2f2c5e",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.0001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_3e9dff76e192bdb82479",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_878848a639e130ad11ad",
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.01,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_ab7b2bbaecc26528fc3a",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.0001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_e313c07c685170294dbd",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_d8544172036bc9ae086b",
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.01,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_bea9b669d36ce0bb0f49",
      "dataset_node_id": "dataset_4d28614559aee4b9",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.0,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_bf1af3070b43afb13889",
      "dataset_node_id": "dataset_4d28614559aee4b9",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.03,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_69c2b002b73676a9e693",
      "dataset_node_id": "dataset_4d28614559aee4b9",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.3,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_bbcaf6772f71b9bfe1cc",
      "dataset_node_id": "dataset_484bdc518daa14bb",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.0,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_decaf5a0597804014f47",
      "dataset_node_id": "dataset_484bdc518daa14bb",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.03,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_5449f78632b714ebdda4",
      "dataset_node_id": "dataset_484bdc518daa14bb",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.3,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_b998d6e03b93bb19946a",
      "dataset_node_id": "dataset_63f24d2a7f7dd6e2",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.0,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_a610f761f13d94f88c47",
      "dataset_node_id": "dataset_63f24d2a7f7dd6e2",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.03,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_13fb3cf7b2dd483e7ad7",
      "dataset_node_id": "dataset_63f24d2a7f7dd6e2",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.3,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_cc8a9186f70e19a657b1",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 16,
        "training_steps": 600,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          150,
          300,
          600
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_f59241db8842e4e80120",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 64,
        "training_steps": 150,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          37,
          75,
          150
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_7a695f3edd302f57ca0e",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 16,
        "training_steps": 600,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          150,
          300,
          600
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_c54739bfbb790f1d5694",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 64,
        "training_steps": 150,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          37,
          75,
          150
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_b08890c32275d0926859",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 16,
        "training_steps": 600,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          150,
          300,
          600
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_df03dcb5bf390a018ea3",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 64,
        "training_steps": 150,
        "total_samples_seen": 9600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          37,
          75,
          150
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_d6d8d7db50c27549e210",
      "dataset_node_id": "dataset_4d28614559aee4b9",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 100,
        "total_samples_seen": 3200,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          25,
          50,
          100
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_87ff1c77bb033dc6bcf6",
      "dataset_node_id": "dataset_4d28614559aee4b9",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 400,
        "total_samples_seen": 12800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          100,
          200,
          400
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_9c4ac004a68ae6f01a23",
      "dataset_node_id": "dataset_4d28614559aee4b9",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_f4212fdb6215bc33b10a",
      "dataset_node_id": "dataset_484bdc518daa14bb",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 100,
        "total_samples_seen": 3200,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          25,
          50,
          100
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_1c8f2761b5434f3d3823",
      "dataset_node_id": "dataset_484bdc518daa14bb",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 400,
        "total_samples_seen": 12800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          100,
          200,
          400
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_24ff1215b04f7b60b52c",
      "dataset_node_id": "dataset_484bdc518daa14bb",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_004d93eda1cef7d7dbf7",
      "dataset_node_id": "dataset_63f24d2a7f7dd6e2",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 100,
        "total_samples_seen": 3200,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          25,
          50,
          100
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_fe3235aaee5ac5aac819",
      "dataset_node_id": "dataset_63f24d2a7f7dd6e2",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 400,
        "total_samples_seen": 12800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          100,
          200,
          400
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_c33cc593b59cf479dc82",
      "dataset_node_id": "dataset_63f24d2a7f7dd6e2",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_1d3aa2831c81134e77af",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 100,
        "total_samples_seen": 3200,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          25,
          50,
          100
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_d8494e4bbb475abc8d6c",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 400,
        "total_samples_seen": 12800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          100,
          200,
          400
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 62284,
      "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91"
    },
    {
      "node_id": "candidate_5ce646d0e89d3b097492",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 800,
        "total_samples_seen": 25600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          200,
          400,
          800
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_fffc5bdace486a294aea",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 100,
        "total_samples_seen": 3200,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          25,
          50,
          100
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 62284,
      "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91"
    },
    {
      "node_id": "candidate_2b14bf0c3a6914ae7431",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 400,
        "total_samples_seen": 12800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          100,
          200,
          400
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_8ad93f31861df974db61",
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 800,
        "total_samples_seen": 25600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          200,
          400,
          800
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 62284,
      "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91"
    },
    {
      "node_id": "candidate_1ee1c85b02c12c69049c",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 100,
        "total_samples_seen": 3200,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          25,
          50,
          100
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_fa921f9d94e2ee00d414",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 400,
        "total_samples_seen": 12800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          100,
          200,
          400
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 62284,
      "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91"
    },
    {
      "node_id": "candidate_5bcc6eb2db18941f07c1",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 800,
        "total_samples_seen": 25600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          200,
          400,
          800
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_a68f85d27604a60fcfac",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 100,
        "total_samples_seen": 3200,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          25,
          50,
          100
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 62284,
      "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91"
    },
    {
      "node_id": "candidate_a9c85c0bf30c5c45b895",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 400,
        "total_samples_seen": 12800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          100,
          200,
          400
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_b0777a15e3036a2b0264",
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 800,
        "total_samples_seen": 25600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          200,
          400,
          800
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 62284,
      "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91"
    },
    {
      "node_id": "candidate_44d23e55659364508d97",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 100,
        "total_samples_seen": 3200,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          25,
          50,
          100
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_dad86d2f0763eb0794d8",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 400,
        "total_samples_seen": 12800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          100,
          200,
          400
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 62284,
      "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91"
    },
    {
      "node_id": "candidate_80324d542043d4a3c3d1",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 800,
        "total_samples_seen": 25600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          200,
          400,
          800
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_72b2c7b6e515710cb184",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 100,
        "total_samples_seen": 3200,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          25,
          50,
          100
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 62284,
      "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91"
    },
    {
      "node_id": "candidate_70fd6488ca27a88be005",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 400,
        "total_samples_seen": 12800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          100,
          200,
          400
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_903f34d3ec924e3bf6a9",
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 7,
        "convolution": "dense",
        "pooling": "spatial",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.001,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 800,
        "total_samples_seen": 25600,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          200,
          400,
          800
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 62284,
      "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91"
    },
    {
      "node_id": "candidate_159df3b6e29ac46bd076",
      "dataset_node_id": "dataset_32df760ef312a38d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "relu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "a93226785f393655cecb8174ae617a22d94e8251a2ae6e3ea53ad51c9e5b4e75"
    },
    {
      "node_id": "candidate_bc1c99efb458f1abecac",
      "dataset_node_id": "dataset_32df760ef312a38d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_f86a78da04bbe1e05364",
      "dataset_node_id": "dataset_32df760ef312a38d",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "tanh",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "101fba5c57246c0fc592fa5b1a30cf7fd569ba0a8001f23a37531234618c32a8"
    },
    {
      "node_id": "candidate_ed8b5935d7cedac78147",
      "dataset_node_id": "dataset_e6c4261755f0ff04",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "relu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "a93226785f393655cecb8174ae617a22d94e8251a2ae6e3ea53ad51c9e5b4e75"
    },
    {
      "node_id": "candidate_e23d523074ffbb679c5c",
      "dataset_node_id": "dataset_e6c4261755f0ff04",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_680d4f9ac1b9852312b9",
      "dataset_node_id": "dataset_e6c4261755f0ff04",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "tanh",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "101fba5c57246c0fc592fa5b1a30cf7fd569ba0a8001f23a37531234618c32a8"
    },
    {
      "node_id": "candidate_f685aa68d7fa486641ee",
      "dataset_node_id": "dataset_220eb81b568bc599",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "relu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "a93226785f393655cecb8174ae617a22d94e8251a2ae6e3ea53ad51c9e5b4e75"
    },
    {
      "node_id": "candidate_f6a336f7bc57e0c58080",
      "dataset_node_id": "dataset_220eb81b568bc599",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f"
    },
    {
      "node_id": "candidate_56051496b194b14c493a",
      "dataset_node_id": "dataset_220eb81b568bc599",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "tanh",
        "residual": true,
        "norm": "layer",
        "norm_first": true,
        "ff_ratio": 2,
        "kernel_size": 3,
        "convolution": "dense",
        "pooling": "mean",
        "pool_grid": 3,
        "coordinates": false,
        "stem_stride": 1,
        "max_parameters": 2000000,
        "input_shape": [
          2,
          24,
          24
        ],
        "output_shape": [
          4
        ],
        "output_kind": "classification"
      },
      "optimizer": {
        "type": "AdamW",
        "lr": 0.003,
        "weight_decay": 0.01,
        "betas": [
          0.9,
          0.999
        ],
        "eps": 1e-08,
        "foreach": false
      },
      "loss": {
        "loss_id": "cross_entropy"
      },
      "training": {
        "batch_size": 32,
        "training_steps": 1200,
        "total_samples_seen": 38400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          300,
          600,
          1200
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 15436,
      "model_code_sha256": "101fba5c57246c0fc592fa5b1a30cf7fd569ba0a8001f23a37531234618c32a8"
    }
  ],
  "model_validation_receipts": [
    {
      "model_sha256": "a499284b051f83f36590fc2c1569f488059fe7224a546edb472973f4297d79a2",
      "parameters": 8572,
      "code_sha256": "2744aac8d669a5a4171096c59dc5d68686ab43164fc80bec2b0f91f9c576457d",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "a32c02e4cadce51c162d745ef287d5b265ca91e78bde84d7ee8d808def586c69",
      "parameters": 31468,
      "code_sha256": "d6ce0d26e9f8f7790ae231b01bdd44e8e90cd75d74db93df02800810ecc27204",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "a1bd473cf79158ecc787efeb9c0dc5134dfd9dedceccb455e6edd07fb87a1784",
      "parameters": 31612,
      "code_sha256": "ddc392f9bdd8df280f3822a219a6653dbbef09d29aa051a52ca5fae1ec2ed0e1",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "38d938395c2b0ec606fe94c997f553a65581c019782d78b1a5fa79b961655450",
      "parameters": 123628,
      "code_sha256": "de12e3d1b8c3c5d6f6b890a1b9e1449f76bb899fd1b6db4d44d794f99ab91150",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "b111d76515a38bc4ff2e978e5ea669d9986a064549d22a050613daf7d4d40e62",
      "parameters": 15436,
      "code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "012ee0ce3318ac0956c6e369d3ecc8ca1e94708e292f17f8a5829777c9431c25",
      "parameters": 15484,
      "code_sha256": "b7f80d8dbdfffa7aca4b82e112df579d6be15d78c5668c7e516661f5532af7d1",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "bc5e6dc614f6120d19c949544c3356b46acd76f61021f322329c99e88df2e734",
      "parameters": 16204,
      "code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "d80d0292ee630eef5226b65f8d2b0f35a0c237eece8bd3b3dec33e463444fc33",
      "parameters": 16252,
      "code_sha256": "f105901133fe76424f8e503f9cc6cca6a84d8017371e00f46f2fde2d43c11841",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "7e4136c390630a90f5dcb4ac7a15764cbdcb80654c3d033e3350e9bd5aa35886",
      "parameters": 17828,
      "code_sha256": "9defa266b5335655ef99314c9606112a78f37e8f7ebf5c14b80373e59e30b747",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "b7507c1223903c04a387a497cad040a98278559e1e3f1e98fb3fe6b4f872e2a4",
      "parameters": 19428,
      "code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "09c94741ce5b01db47a4436d4397047bed7c4390a3390be9cee94865f382d210",
      "parameters": 20068,
      "code_sha256": "e71b2de720715f358b0a505a393993f429c24e76a254225b74c00743a9c38227",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "deb6553fcc5c72f5e0fc6c206ce86f20fbc276715606b12b5e5ee9ba9371562d",
      "parameters": 19876,
      "code_sha256": "f29b6e6a38c3792bb771f0a25777e143f7555c8c6cf80cd1cd0007277960f5b3",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "9741680be227696a95ed50bd5cc4ef7b714370ecbb5434dc6663b6e83eaa8101",
      "parameters": 18276,
      "code_sha256": "fbe5f23efc3b8d68938d38f8539601ea9f2b82f6121f23d6f218d9256e9bb082",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "858e0159e017e628901adb1dceed7df5ca58291db4964633434a2b185631dd1e",
      "parameters": 19556,
      "code_sha256": "13e3d8cca256e65b2821e79f624f79f2ea20c5314d05efe3f9e68aa9b4913a35",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "1ac2ee9db39b9fabf45ca3f5351e022e645806f176f63350feac0b76673c1a0a",
      "parameters": 129700,
      "code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "137bd0fc1cac08dbf41194971231e180a55c4cf433e7eb0f7abbf9fdb27ede7d",
      "parameters": 34012,
      "code_sha256": "394a7050f24d49bb1c1fe5a84fddde7fb349d74bb928ebd3d56918f08922ce9e",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "c4d85b46664ffc54079d35f9115e00b5beafae685260aa578d7427db062cf164",
      "parameters": 16156,
      "code_sha256": "247e1e707578507112126325851aaa80fe6e7e885427ce689c887595d7b6e215",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "32c0dee4342942266a24ebde05394e5e9184ea14dbc5f0603cb212d0311bdaa1",
      "parameters": 33868,
      "code_sha256": "e9a9a2765f20d57dccca28a3fffdc7dadefe39637b32ed6013d958be3159f651",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "7d0a78936209bbffa55a806b041f61ae61199aeeebebaf9d7635df2a1a199e34",
      "parameters": 15580,
      "code_sha256": "4eed0b51c7ee7ec1f74effe0bdf240b1ba24009b9af627dd4c278268ccdac078",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "dd0477c9358b5b307ff5e4a250f890465a8019f877a6b22d1afb1fe9001f6316",
      "parameters": 34588,
      "code_sha256": "df29ae51767411dcd43886969870251fe836bfde85d67f093e5d3a629a91112d",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "ad7553929580de63afa3ca057c55e520f91aba217a7209ea3e0afc847e0d9732",
      "parameters": 5500,
      "code_sha256": "297bdbdefb2459efc98a425849209a52ad5c0084bb3bc416d73fb096109e7e4e",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "03c4f8058cc0963007813127aeb54a26e4315ae9e98cea3540a992467b97c04b",
      "parameters": 20212,
      "code_sha256": "c5b04c3c387f9d97858c89cc93dd742f98b4fc635d9ddef274224866d4a35f0a",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "9e85a417345966baa28d046ee501549fb02bd2699e3d8c70dd8694eda8a9c049",
      "parameters": 60820,
      "code_sha256": "751ce9f6d2824bfbe26fced8a343596c9f2f627a889f39ed9cd667537564058c",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "55908d054620dbe31b32e50f1b6cfcd89dc117993fe549fe3e4d399bf37c9023",
      "parameters": 16204,
      "code_sha256": "487f99a81970de56cd83e5ed249732dbfa7860e0a5389abfec6823e543153207",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "8f4239a502c5285d708f2eab56516d97da9a365eb20637d80e49988b8952a944",
      "parameters": 46732,
      "code_sha256": "7183d61f2bf2188b283e14ce82318d20f5855fcd0a16bc4a57456c2081c89b67",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "6617ecba02609c502d839d6a8f62c5923d9cc5fc4ca6cced1a19cbef04fed272",
      "parameters": 46732,
      "code_sha256": "7818c319584d6ddd071c6fa738ca50c22811f9f460d2e6e4469a54be52606813",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "977f319e0cea3dcc2cd0086e5ed2af49705af1cc2b86fab654dd19facc8a45a1",
      "parameters": 15340,
      "code_sha256": "384ea3383ad48c29b344417d8d4ed4a0692433ae1e3d913dceeb3f2764b1ed74",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "25de93201e8c0441896db81f1c5aa72b4a75a141a44abe31f2c76999ef647a13",
      "parameters": 15436,
      "code_sha256": "c08a63248209b0d5a64c9236d54c5be7f1beb598dc26e8c73271614ded80629b",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "d464dd82e29272236f813106b00fcb0c7b17a5b391ee1fea927957811fe64c20",
      "parameters": 15340,
      "code_sha256": "ea5fa570ff31093c95f665efc19d7c1108483d3113df92e6417e1a5f1fe6a40e",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "ce71a31b71a57a0f014c97132a061f746d87d3db9edca1e83d989124b41a0bfe",
      "parameters": 13084,
      "code_sha256": "d9997a6866de1641a44db88c5d3ea82cbc449f26a4ffc4d05ac1a573bbcb0a30",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "5764b6f35b9b8627720555e932ce04347f9bc2b45d3de0fd98089e2ab5c3a939",
      "parameters": 20140,
      "code_sha256": "c3cac7123a9b71d029392ad0bd01da692d128efe5f3f2267300a9d4cd61e5659",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "7b1f44c0e0e19e83c60dcdbb326fdab7455fcbec559fe864d527d712905c2f6e",
      "parameters": 13084,
      "code_sha256": "1486f09e036eb722ea94783ce64d303a75e37f0d35a5b0010bf6d3362b2bb66c",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "e5e3305a872fad48d94315c75bfbf2731923df8b9b81e83e86509e8f18a8074e",
      "parameters": 15436,
      "code_sha256": "a93226785f393655cecb8174ae617a22d94e8251a2ae6e3ea53ad51c9e5b4e75",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "8216f5d64571fb72df727b76a5f9cf54c09afa62c1996bc0b41682914a21d67d",
      "parameters": 20140,
      "code_sha256": "32cca09dafc2c833b1906f5574e0c99166b7f54576f13dae140418f49f45a2cb",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "a5df575c5cf07df89142f4d4a5064d7191115c9426244de107ab84a072f10c1d",
      "parameters": 27748,
      "code_sha256": "33aa4418bafc50c1f23adb582ed959f04f2cfe8daf97db6cd775f669564215b7",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "4f8eb4de41e91ad507b92b8a966a7369acaa72b85e24e78333053928f08c4081",
      "parameters": 15268,
      "code_sha256": "37b0b123040119e40aa10bdfce698d951381255500bad19d4d734cb820a3bc1d",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "86bc4ef668449524b9513c281561b94a0431a28043455d269cb721d824508229",
      "parameters": 19428,
      "code_sha256": "4038255caa48fa6b4fa7829427ab55fcf105ad98de0ac407c33ee34def5a043a",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "3eb01f4879f354df32b6c2eb77826abf6f6a2a7721d967fd91dc067d3e65ec9e",
      "parameters": 19428,
      "code_sha256": "9e76b05e6a7a0f7d75de4c9fd9ac4652a7ce37e4f12081897317f660e7ca7621",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "889b23506525aed971917e7814a606756db6a68623844ce02b649477bbef4122",
      "parameters": 15268,
      "code_sha256": "8364bbae844edba18d6067aa23e9bf7faaded3ed3722211383bbd3e63c55c6da",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "a17e477e10a597a6ff092efd8eaa665d14f383e32545cbc3c43784b0dadd9e9c",
      "parameters": 27748,
      "code_sha256": "5e4a2c32ed8426dea1fabecda2b813b8ff8e4a7d0fcafbd8be819520e625e953",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "c0c9b8171f60513e8b6564a43ad8e00fd6446e1281259e5ff1cfa132d1092916",
      "parameters": 62284,
      "code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "83b63ada8bd4faf8fe40201ac47492ae6b4f816a9b8388d99e6b4f9d2fc6ff6d",
      "parameters": 15436,
      "code_sha256": "101fba5c57246c0fc592fa5b1a30cf7fd569ba0a8001f23a37531234618c32a8",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    }
  ],
  "templates": [
    {
      "id": "receptive_field_depth_s841901",
      "category": "spatial_support",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.kernel_size",
        "model.depth"
      ],
      "hypothesis": "Kernel support and depth interact under finite optimization.",
      "explanation_target": "Use the full 2x2 kernel/depth control; spatial pooling itself preserves coarse positions, so kernel support is not the sole available route.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "receptive_field_depth",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.depth",
          "model.kernel_size"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "k3_d1": {
            "model.depth": 1,
            "model.kernel_size": 3
          },
          "k3_d4": {
            "model.depth": 4,
            "model.kernel_size": 3
          },
          "k7_d1": {
            "model.depth": 1,
            "model.kernel_size": 7
          }
        },
        "actual_changed_fields": [
          "model.depth",
          "model.kernel_size"
        ]
      },
      "choices": [
        {
          "id": "k3_d1",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 1,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 8572,
          "model_code_sha256": "2744aac8d669a5a4171096c59dc5d68686ab43164fc80bec2b0f91f9c576457d",
          "node_id": "candidate_aedc559262a997543320"
        },
        {
          "id": "k3_d4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 4,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 31468,
          "model_code_sha256": "d6ce0d26e9f8f7790ae231b01bdd44e8e90cd75d74db93df02800810ecc27204",
          "node_id": "candidate_e2e98633894e1a273735"
        },
        {
          "id": "k7_d1",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 1,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 31612,
          "model_code_sha256": "ddc392f9bdd8df280f3822a219a6653dbbef09d29aa051a52ca5fae1ec2ed0e1",
          "node_id": "candidate_bb23efca29cbc84e0f54"
        }
      ],
      "diagnostics": [
        {
          "id": "k7_d4",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 4,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 123628,
          "model_code_sha256": "de12e3d1b8c3c5d6f6b890a1b9e1449f76bb899fd1b6db4d44d794f99ab91150",
          "node_id": "candidate_96057a489f621815967a"
        }
      ]
    },
    {
      "id": "receptive_field_depth_s841902",
      "category": "spatial_support",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.kernel_size",
        "model.depth"
      ],
      "hypothesis": "Kernel support and depth interact under finite optimization.",
      "explanation_target": "Use the full 2x2 kernel/depth control; spatial pooling itself preserves coarse positions, so kernel support is not the sole available route.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "receptive_field_depth",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.depth",
          "model.kernel_size"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "k3_d1": {
            "model.depth": 1,
            "model.kernel_size": 3
          },
          "k3_d4": {
            "model.depth": 4,
            "model.kernel_size": 3
          },
          "k7_d1": {
            "model.depth": 1,
            "model.kernel_size": 7
          }
        },
        "actual_changed_fields": [
          "model.depth",
          "model.kernel_size"
        ]
      },
      "choices": [
        {
          "id": "k3_d1",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 1,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 8572,
          "model_code_sha256": "2744aac8d669a5a4171096c59dc5d68686ab43164fc80bec2b0f91f9c576457d",
          "node_id": "candidate_5c73c2b705cd3ac05e84"
        },
        {
          "id": "k3_d4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 4,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 31468,
          "model_code_sha256": "d6ce0d26e9f8f7790ae231b01bdd44e8e90cd75d74db93df02800810ecc27204",
          "node_id": "candidate_82928db599c08a734f6d"
        },
        {
          "id": "k7_d1",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 1,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 31612,
          "model_code_sha256": "ddc392f9bdd8df280f3822a219a6653dbbef09d29aa051a52ca5fae1ec2ed0e1",
          "node_id": "candidate_264ad6f33a2d476f3ffd"
        }
      ],
      "diagnostics": [
        {
          "id": "k7_d4",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 4,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 123628,
          "model_code_sha256": "de12e3d1b8c3c5d6f6b890a1b9e1449f76bb899fd1b6db4d44d794f99ab91150",
          "node_id": "candidate_c241ce5e2ffc71265de4"
        }
      ]
    },
    {
      "id": "receptive_field_depth_s841903",
      "category": "spatial_support",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.kernel_size",
        "model.depth"
      ],
      "hypothesis": "Kernel support and depth interact under finite optimization.",
      "explanation_target": "Use the full 2x2 kernel/depth control; spatial pooling itself preserves coarse positions, so kernel support is not the sole available route.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "receptive_field_depth",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.depth",
          "model.kernel_size"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "k3_d1": {
            "model.depth": 1,
            "model.kernel_size": 3
          },
          "k3_d4": {
            "model.depth": 4,
            "model.kernel_size": 3
          },
          "k7_d1": {
            "model.depth": 1,
            "model.kernel_size": 7
          }
        },
        "actual_changed_fields": [
          "model.depth",
          "model.kernel_size"
        ]
      },
      "choices": [
        {
          "id": "k3_d1",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 1,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 8572,
          "model_code_sha256": "2744aac8d669a5a4171096c59dc5d68686ab43164fc80bec2b0f91f9c576457d",
          "node_id": "candidate_016ee74c9be9565d2e2f"
        },
        {
          "id": "k3_d4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 4,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 31468,
          "model_code_sha256": "d6ce0d26e9f8f7790ae231b01bdd44e8e90cd75d74db93df02800810ecc27204",
          "node_id": "candidate_0e6bb18a73baeeb2cd6e"
        },
        {
          "id": "k7_d1",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 1,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 31612,
          "model_code_sha256": "ddc392f9bdd8df280f3822a219a6653dbbef09d29aa051a52ca5fae1ec2ed0e1",
          "node_id": "candidate_be9ddfc1b3d86605ec66"
        }
      ],
      "diagnostics": [
        {
          "id": "k7_d4",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 4,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 123628,
          "model_code_sha256": "de12e3d1b8c3c5d6f6b890a1b9e1449f76bb899fd1b6db4d44d794f99ab91150",
          "node_id": "candidate_d8dfd33851eb7eb58bf5"
        }
      ]
    },
    {
      "id": "pooling_coordinate_s841901",
      "category": "spatial_readout",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.pooling",
        "model.coordinates"
      ],
      "hypothesis": "Coordinate input and spatial readout provide different location information.",
      "explanation_target": "Use the completed 2x2 control and distinguish representation availability from optimization; convolution boundary effects mean finite images are not exactly translation invariant.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "pooling_coordinate",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.coordinates",
          "model.pooling"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "mean": {
            "model.coordinates": false,
            "model.pooling": "mean"
          },
          "coordinate_mean": {
            "model.coordinates": true,
            "model.pooling": "mean"
          },
          "spatial": {
            "model.coordinates": false,
            "model.pooling": "spatial"
          }
        },
        "actual_changed_fields": [
          "model.coordinates",
          "model.pooling"
        ]
      },
      "choices": [
        {
          "id": "mean",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_3cbaed19cf1f1bd78892"
        },
        {
          "id": "coordinate_mean",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": true,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15484,
          "model_code_sha256": "b7f80d8dbdfffa7aca4b82e112df579d6be15d78c5668c7e516661f5532af7d1",
          "node_id": "candidate_fe248f3711cecdf66c13"
        },
        {
          "id": "spatial",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_675b0f0945a65c7728d6"
        }
      ],
      "diagnostics": [
        {
          "id": "coordinate_spatial",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": true,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16252,
          "model_code_sha256": "f105901133fe76424f8e503f9cc6cca6a84d8017371e00f46f2fde2d43c11841",
          "node_id": "candidate_4e616725a8ecc13c433b"
        }
      ]
    },
    {
      "id": "pooling_coordinate_s841902",
      "category": "spatial_readout",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.pooling",
        "model.coordinates"
      ],
      "hypothesis": "Coordinate input and spatial readout provide different location information.",
      "explanation_target": "Use the completed 2x2 control and distinguish representation availability from optimization; convolution boundary effects mean finite images are not exactly translation invariant.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "pooling_coordinate",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.coordinates",
          "model.pooling"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "mean": {
            "model.coordinates": false,
            "model.pooling": "mean"
          },
          "coordinate_mean": {
            "model.coordinates": true,
            "model.pooling": "mean"
          },
          "spatial": {
            "model.coordinates": false,
            "model.pooling": "spatial"
          }
        },
        "actual_changed_fields": [
          "model.coordinates",
          "model.pooling"
        ]
      },
      "choices": [
        {
          "id": "mean",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_c93709ee26aacbc9f437"
        },
        {
          "id": "coordinate_mean",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": true,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15484,
          "model_code_sha256": "b7f80d8dbdfffa7aca4b82e112df579d6be15d78c5668c7e516661f5532af7d1",
          "node_id": "candidate_d91d654be99bc328c040"
        },
        {
          "id": "spatial",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_f9bdfa11ef52bc5e3528"
        }
      ],
      "diagnostics": [
        {
          "id": "coordinate_spatial",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": true,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16252,
          "model_code_sha256": "f105901133fe76424f8e503f9cc6cca6a84d8017371e00f46f2fde2d43c11841",
          "node_id": "candidate_a644642ee0d58a7cbf09"
        }
      ]
    },
    {
      "id": "pooling_coordinate_s841903",
      "category": "spatial_readout",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.pooling",
        "model.coordinates"
      ],
      "hypothesis": "Coordinate input and spatial readout provide different location information.",
      "explanation_target": "Use the completed 2x2 control and distinguish representation availability from optimization; convolution boundary effects mean finite images are not exactly translation invariant.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "pooling_coordinate",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.coordinates",
          "model.pooling"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "mean": {
            "model.coordinates": false,
            "model.pooling": "mean"
          },
          "coordinate_mean": {
            "model.coordinates": true,
            "model.pooling": "mean"
          },
          "spatial": {
            "model.coordinates": false,
            "model.pooling": "spatial"
          }
        },
        "actual_changed_fields": [
          "model.coordinates",
          "model.pooling"
        ]
      },
      "choices": [
        {
          "id": "mean",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_07cf9b9e2ef0417e1afc"
        },
        {
          "id": "coordinate_mean",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": true,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15484,
          "model_code_sha256": "b7f80d8dbdfffa7aca4b82e112df579d6be15d78c5668c7e516661f5532af7d1",
          "node_id": "candidate_c65bf21470c8b1e81842"
        },
        {
          "id": "spatial",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_d6f536a09652df88c280"
        }
      ],
      "diagnostics": [
        {
          "id": "coordinate_spatial",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": true,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16252,
          "model_code_sha256": "f105901133fe76424f8e503f9cc6cca6a84d8017371e00f46f2fde2d43c11841",
          "node_id": "candidate_4b043b424317b46e875a"
        }
      ]
    },
    {
      "id": "patch_position_s841901",
      "category": "patch",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.patch_size",
        "model.position"
      ],
      "hypothesis": "Token scale and learned position information can trade off at fixed steps.",
      "explanation_target": "All 3x2 patch/position cells are available. Patchification preserves within-patch arrangement even without positional embeddings.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "patch_position",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.patch_size",
          "model.position"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "p3_none": {
            "model.patch_size": 3,
            "model.position": "none"
          },
          "p4_learned": {
            "model.patch_size": 4,
            "model.position": "learned"
          },
          "p6_learned": {
            "model.patch_size": 6,
            "model.position": "learned"
          }
        },
        "actual_changed_fields": [
          "model.patch_size",
          "model.position"
        ]
      },
      "choices": [
        {
          "id": "p3_none",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 3,
            "position": "none",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 17828,
          "model_code_sha256": "9defa266b5335655ef99314c9606112a78f37e8f7ebf5c14b80373e59e30b747",
          "node_id": "candidate_cb13dd586f5542e7de21"
        },
        {
          "id": "p4_learned",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_5cf50db65aa67995a4f9"
        },
        {
          "id": "p6_learned",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 6,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20068,
          "model_code_sha256": "e71b2de720715f358b0a505a393993f429c24e76a254225b74c00743a9c38227",
          "node_id": "candidate_f40fb6fc801f6d2a3c95"
        }
      ],
      "diagnostics": [
        {
          "id": "p3_learned",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 3,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19876,
          "model_code_sha256": "f29b6e6a38c3792bb771f0a25777e143f7555c8c6cf80cd1cd0007277960f5b3",
          "node_id": "candidate_e7f426dc4d829a438fe9"
        },
        {
          "id": "p4_none",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "none",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 18276,
          "model_code_sha256": "fbe5f23efc3b8d68938d38f8539601ea9f2b82f6121f23d6f218d9256e9bb082",
          "node_id": "candidate_159ce277330e0990eb83"
        },
        {
          "id": "p6_none",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 6,
            "position": "none",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19556,
          "model_code_sha256": "13e3d8cca256e65b2821e79f624f79f2ea20c5314d05efe3f9e68aa9b4913a35",
          "node_id": "candidate_970c26429b51319c117b"
        }
      ]
    },
    {
      "id": "patch_position_s841902",
      "category": "patch",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.patch_size",
        "model.position"
      ],
      "hypothesis": "Token scale and learned position information can trade off at fixed steps.",
      "explanation_target": "All 3x2 patch/position cells are available. Patchification preserves within-patch arrangement even without positional embeddings.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "patch_position",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.patch_size",
          "model.position"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "p3_none": {
            "model.patch_size": 3,
            "model.position": "none"
          },
          "p4_learned": {
            "model.patch_size": 4,
            "model.position": "learned"
          },
          "p6_learned": {
            "model.patch_size": 6,
            "model.position": "learned"
          }
        },
        "actual_changed_fields": [
          "model.patch_size",
          "model.position"
        ]
      },
      "choices": [
        {
          "id": "p3_none",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 3,
            "position": "none",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 17828,
          "model_code_sha256": "9defa266b5335655ef99314c9606112a78f37e8f7ebf5c14b80373e59e30b747",
          "node_id": "candidate_9096f1890ddea0000c15"
        },
        {
          "id": "p4_learned",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_5313f05220195b0f67e9"
        },
        {
          "id": "p6_learned",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 6,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20068,
          "model_code_sha256": "e71b2de720715f358b0a505a393993f429c24e76a254225b74c00743a9c38227",
          "node_id": "candidate_76dacecff1eed9e2c8e8"
        }
      ],
      "diagnostics": [
        {
          "id": "p3_learned",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 3,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19876,
          "model_code_sha256": "f29b6e6a38c3792bb771f0a25777e143f7555c8c6cf80cd1cd0007277960f5b3",
          "node_id": "candidate_7cb8c5ca7cbb9b8fd732"
        },
        {
          "id": "p4_none",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "none",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 18276,
          "model_code_sha256": "fbe5f23efc3b8d68938d38f8539601ea9f2b82f6121f23d6f218d9256e9bb082",
          "node_id": "candidate_5314c84c867a5b81a5ea"
        },
        {
          "id": "p6_none",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 6,
            "position": "none",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19556,
          "model_code_sha256": "13e3d8cca256e65b2821e79f624f79f2ea20c5314d05efe3f9e68aa9b4913a35",
          "node_id": "candidate_41c996694bfdeb01d4eb"
        }
      ]
    },
    {
      "id": "patch_position_s841903",
      "category": "patch",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.patch_size",
        "model.position"
      ],
      "hypothesis": "Token scale and learned position information can trade off at fixed steps.",
      "explanation_target": "All 3x2 patch/position cells are available. Patchification preserves within-patch arrangement even without positional embeddings.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "patch_position",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.patch_size",
          "model.position"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "p3_none": {
            "model.patch_size": 3,
            "model.position": "none"
          },
          "p4_learned": {
            "model.patch_size": 4,
            "model.position": "learned"
          },
          "p6_learned": {
            "model.patch_size": 6,
            "model.position": "learned"
          }
        },
        "actual_changed_fields": [
          "model.patch_size",
          "model.position"
        ]
      },
      "choices": [
        {
          "id": "p3_none",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 3,
            "position": "none",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 17828,
          "model_code_sha256": "9defa266b5335655ef99314c9606112a78f37e8f7ebf5c14b80373e59e30b747",
          "node_id": "candidate_365c90b1233692d34e21"
        },
        {
          "id": "p4_learned",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_681cbfb92210830d5199"
        },
        {
          "id": "p6_learned",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 6,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20068,
          "model_code_sha256": "e71b2de720715f358b0a505a393993f429c24e76a254225b74c00743a9c38227",
          "node_id": "candidate_dfb5edbd1f2cbf291954"
        }
      ],
      "diagnostics": [
        {
          "id": "p3_learned",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 3,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19876,
          "model_code_sha256": "f29b6e6a38c3792bb771f0a25777e143f7555c8c6cf80cd1cd0007277960f5b3",
          "node_id": "candidate_98e22078313daa99d5d2"
        },
        {
          "id": "p4_none",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "none",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 18276,
          "model_code_sha256": "fbe5f23efc3b8d68938d38f8539601ea9f2b82f6121f23d6f218d9256e9bb082",
          "node_id": "candidate_8c676e89972ddefe6f92"
        },
        {
          "id": "p6_none",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 6,
            "position": "none",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19556,
          "model_code_sha256": "13e3d8cca256e65b2821e79f624f79f2ea20c5314d05efe3f9e68aa9b4913a35",
          "node_id": "candidate_4866c13502bb9e784e3e"
        }
      ]
    },
    {
      "id": "operator_relation_s841901",
      "category": "operator",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.mode"
      ],
      "hypothesis": "Three different inductive biases solve the same relational image task.",
      "explanation_target": "This is a bundled architecture recipe comparison with actual changed fields and parameter counts; do not attribute the entire gap to one named operator.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "operator_relation",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.mode",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width"
        ],
        "undeclared_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width"
        ],
        "comparison_class": "bundled_recipe",
        "field_values": {
          "cnn": {
            "model.convolution": "dense",
            "model.coordinates": false,
            "model.ff_ratio": 2,
            "model.heads": "<absent>",
            "model.kernel_size": 3,
            "model.mode": "cnn",
            "model.norm": "layer",
            "model.norm_first": true,
            "model.patch_size": "<absent>",
            "model.pool_grid": 3,
            "model.pooling": "spatial",
            "model.position": "<absent>",
            "model.stem_stride": 1,
            "model.width": 24
          },
          "flat": {
            "model.convolution": "<absent>",
            "model.coordinates": "<absent>",
            "model.ff_ratio": "<absent>",
            "model.heads": "<absent>",
            "model.kernel_size": "<absent>",
            "model.mode": "mlp",
            "model.norm": "<absent>",
            "model.norm_first": "<absent>",
            "model.patch_size": "<absent>",
            "model.pool_grid": "<absent>",
            "model.pooling": "<absent>",
            "model.position": "<absent>",
            "model.stem_stride": "<absent>",
            "model.width": 96
          },
          "attention": {
            "model.convolution": "<absent>",
            "model.coordinates": "<absent>",
            "model.ff_ratio": 2,
            "model.heads": 4,
            "model.kernel_size": "<absent>",
            "model.mode": "patch_transformer",
            "model.norm": "layer",
            "model.norm_first": true,
            "model.patch_size": 4,
            "model.pool_grid": "<absent>",
            "model.pooling": "<absent>",
            "model.position": "learned",
            "model.stem_stride": "<absent>",
            "model.width": 32
          }
        },
        "actual_changed_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.mode",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width"
        ]
      },
      "choices": [
        {
          "id": "cnn",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_675b0f0945a65c7728d6"
        },
        {
          "id": "flat",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "mlp",
            "width": 96,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 129700,
          "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010",
          "node_id": "candidate_9fe585c7b9d737d48244"
        },
        {
          "id": "attention",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_5cf50db65aa67995a4f9"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "operator_relation_s841902",
      "category": "operator",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.mode"
      ],
      "hypothesis": "Three different inductive biases solve the same relational image task.",
      "explanation_target": "This is a bundled architecture recipe comparison with actual changed fields and parameter counts; do not attribute the entire gap to one named operator.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "operator_relation",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.mode",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width"
        ],
        "undeclared_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width"
        ],
        "comparison_class": "bundled_recipe",
        "field_values": {
          "cnn": {
            "model.convolution": "dense",
            "model.coordinates": false,
            "model.ff_ratio": 2,
            "model.heads": "<absent>",
            "model.kernel_size": 3,
            "model.mode": "cnn",
            "model.norm": "layer",
            "model.norm_first": true,
            "model.patch_size": "<absent>",
            "model.pool_grid": 3,
            "model.pooling": "spatial",
            "model.position": "<absent>",
            "model.stem_stride": 1,
            "model.width": 24
          },
          "flat": {
            "model.convolution": "<absent>",
            "model.coordinates": "<absent>",
            "model.ff_ratio": "<absent>",
            "model.heads": "<absent>",
            "model.kernel_size": "<absent>",
            "model.mode": "mlp",
            "model.norm": "<absent>",
            "model.norm_first": "<absent>",
            "model.patch_size": "<absent>",
            "model.pool_grid": "<absent>",
            "model.pooling": "<absent>",
            "model.position": "<absent>",
            "model.stem_stride": "<absent>",
            "model.width": 96
          },
          "attention": {
            "model.convolution": "<absent>",
            "model.coordinates": "<absent>",
            "model.ff_ratio": 2,
            "model.heads": 4,
            "model.kernel_size": "<absent>",
            "model.mode": "patch_transformer",
            "model.norm": "layer",
            "model.norm_first": true,
            "model.patch_size": 4,
            "model.pool_grid": "<absent>",
            "model.pooling": "<absent>",
            "model.position": "learned",
            "model.stem_stride": "<absent>",
            "model.width": 32
          }
        },
        "actual_changed_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.mode",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width"
        ]
      },
      "choices": [
        {
          "id": "cnn",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_f9bdfa11ef52bc5e3528"
        },
        {
          "id": "flat",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "mlp",
            "width": 96,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 129700,
          "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010",
          "node_id": "candidate_eb0fc9a5ce2767f19d93"
        },
        {
          "id": "attention",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_5313f05220195b0f67e9"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "operator_relation_s841903",
      "category": "operator",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.mode"
      ],
      "hypothesis": "Three different inductive biases solve the same relational image task.",
      "explanation_target": "This is a bundled architecture recipe comparison with actual changed fields and parameter counts; do not attribute the entire gap to one named operator.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "operator_relation",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.mode",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width"
        ],
        "undeclared_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width"
        ],
        "comparison_class": "bundled_recipe",
        "field_values": {
          "cnn": {
            "model.convolution": "dense",
            "model.coordinates": false,
            "model.ff_ratio": 2,
            "model.heads": "<absent>",
            "model.kernel_size": 3,
            "model.mode": "cnn",
            "model.norm": "layer",
            "model.norm_first": true,
            "model.patch_size": "<absent>",
            "model.pool_grid": 3,
            "model.pooling": "spatial",
            "model.position": "<absent>",
            "model.stem_stride": 1,
            "model.width": 24
          },
          "flat": {
            "model.convolution": "<absent>",
            "model.coordinates": "<absent>",
            "model.ff_ratio": "<absent>",
            "model.heads": "<absent>",
            "model.kernel_size": "<absent>",
            "model.mode": "mlp",
            "model.norm": "<absent>",
            "model.norm_first": "<absent>",
            "model.patch_size": "<absent>",
            "model.pool_grid": "<absent>",
            "model.pooling": "<absent>",
            "model.position": "<absent>",
            "model.stem_stride": "<absent>",
            "model.width": 96
          },
          "attention": {
            "model.convolution": "<absent>",
            "model.coordinates": "<absent>",
            "model.ff_ratio": 2,
            "model.heads": 4,
            "model.kernel_size": "<absent>",
            "model.mode": "patch_transformer",
            "model.norm": "layer",
            "model.norm_first": true,
            "model.patch_size": 4,
            "model.pool_grid": "<absent>",
            "model.pooling": "<absent>",
            "model.position": "learned",
            "model.stem_stride": "<absent>",
            "model.width": 32
          }
        },
        "actual_changed_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.mode",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width"
        ]
      },
      "choices": [
        {
          "id": "cnn",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_d6f536a09652df88c280"
        },
        {
          "id": "flat",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "mlp",
            "width": 96,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 129700,
          "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010",
          "node_id": "candidate_32c63bc42039bb009be4"
        },
        {
          "id": "attention",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_681cbfb92210830d5199"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "operator_local_shape_s841901",
      "category": "operator",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.mode"
      ],
      "hypothesis": "Local shape recognition need not favor the same recipe as relative positions.",
      "explanation_target": "Compare operator recipes within this fixed task; the CNN readout control distinguishes mean pooling from changing the operator.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "operator_local_shape",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_68a832a939c0b770",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.mode",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width",
          "optimizer.lr"
        ],
        "undeclared_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width",
          "optimizer.lr"
        ],
        "comparison_class": "bundled_recipe",
        "field_values": {
          "cnn_mean": {
            "model.convolution": "dense",
            "model.coordinates": false,
            "model.ff_ratio": 2,
            "model.heads": "<absent>",
            "model.kernel_size": 3,
            "model.mode": "cnn",
            "model.norm": "layer",
            "model.norm_first": true,
            "model.patch_size": "<absent>",
            "model.pool_grid": 3,
            "model.pooling": "mean",
            "model.position": "<absent>",
            "model.stem_stride": 1,
            "model.width": 24,
            "optimizer.lr": 0.003
          },
          "flat": {
            "model.convolution": "<absent>",
            "model.coordinates": "<absent>",
            "model.ff_ratio": "<absent>",
            "model.heads": "<absent>",
            "model.kernel_size": "<absent>",
            "model.mode": "mlp",
            "model.norm": "<absent>",
            "model.norm_first": "<absent>",
            "model.patch_size": "<absent>",
            "model.pool_grid": "<absent>",
            "model.pooling": "<absent>",
            "model.position": "<absent>",
            "model.stem_stride": "<absent>",
            "model.width": 96,
            "optimizer.lr": 0.001
          },
          "attention": {
            "model.convolution": "<absent>",
            "model.coordinates": "<absent>",
            "model.ff_ratio": 2,
            "model.heads": 4,
            "model.kernel_size": "<absent>",
            "model.mode": "patch_transformer",
            "model.norm": "layer",
            "model.norm_first": true,
            "model.patch_size": 4,
            "model.pool_grid": "<absent>",
            "model.pooling": "<absent>",
            "model.position": "learned",
            "model.stem_stride": "<absent>",
            "model.width": 32,
            "optimizer.lr": 0.001
          }
        },
        "actual_changed_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.mode",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width",
          "optimizer.lr"
        ]
      },
      "choices": [
        {
          "id": "cnn_mean",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_3b74befeb2c6c9eb61c9"
        },
        {
          "id": "flat",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "mlp",
            "width": 96,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 129700,
          "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010",
          "node_id": "candidate_5387a55740ddb1ad3894"
        },
        {
          "id": "attention",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_ac04e963dc5200162093"
        }
      ],
      "diagnostics": [
        {
          "id": "cnn_spatial",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_c5a024c8f4d8af48cf51"
        }
      ]
    },
    {
      "id": "operator_local_shape_s841902",
      "category": "operator",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.mode"
      ],
      "hypothesis": "Local shape recognition need not favor the same recipe as relative positions.",
      "explanation_target": "Compare operator recipes within this fixed task; the CNN readout control distinguishes mean pooling from changing the operator.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "operator_local_shape",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.mode",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width",
          "optimizer.lr"
        ],
        "undeclared_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width",
          "optimizer.lr"
        ],
        "comparison_class": "bundled_recipe",
        "field_values": {
          "cnn_mean": {
            "model.convolution": "dense",
            "model.coordinates": false,
            "model.ff_ratio": 2,
            "model.heads": "<absent>",
            "model.kernel_size": 3,
            "model.mode": "cnn",
            "model.norm": "layer",
            "model.norm_first": true,
            "model.patch_size": "<absent>",
            "model.pool_grid": 3,
            "model.pooling": "mean",
            "model.position": "<absent>",
            "model.stem_stride": 1,
            "model.width": 24,
            "optimizer.lr": 0.003
          },
          "flat": {
            "model.convolution": "<absent>",
            "model.coordinates": "<absent>",
            "model.ff_ratio": "<absent>",
            "model.heads": "<absent>",
            "model.kernel_size": "<absent>",
            "model.mode": "mlp",
            "model.norm": "<absent>",
            "model.norm_first": "<absent>",
            "model.patch_size": "<absent>",
            "model.pool_grid": "<absent>",
            "model.pooling": "<absent>",
            "model.position": "<absent>",
            "model.stem_stride": "<absent>",
            "model.width": 96,
            "optimizer.lr": 0.001
          },
          "attention": {
            "model.convolution": "<absent>",
            "model.coordinates": "<absent>",
            "model.ff_ratio": 2,
            "model.heads": 4,
            "model.kernel_size": "<absent>",
            "model.mode": "patch_transformer",
            "model.norm": "layer",
            "model.norm_first": true,
            "model.patch_size": 4,
            "model.pool_grid": "<absent>",
            "model.pooling": "<absent>",
            "model.position": "learned",
            "model.stem_stride": "<absent>",
            "model.width": 32,
            "optimizer.lr": 0.001
          }
        },
        "actual_changed_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.mode",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width",
          "optimizer.lr"
        ]
      },
      "choices": [
        {
          "id": "cnn_mean",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_7399cba3f8ca90f400cc"
        },
        {
          "id": "flat",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "mlp",
            "width": 96,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 129700,
          "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010",
          "node_id": "candidate_243be9f0f4d2817e59a6"
        },
        {
          "id": "attention",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_6cb11cedc2472110e7b0"
        }
      ],
      "diagnostics": [
        {
          "id": "cnn_spatial",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_c79b5e3e8168a286054c"
        }
      ]
    },
    {
      "id": "operator_local_shape_s841903",
      "category": "operator",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.mode"
      ],
      "hypothesis": "Local shape recognition need not favor the same recipe as relative positions.",
      "explanation_target": "Compare operator recipes within this fixed task; the CNN readout control distinguishes mean pooling from changing the operator.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "operator_local_shape",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.mode",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width",
          "optimizer.lr"
        ],
        "undeclared_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width",
          "optimizer.lr"
        ],
        "comparison_class": "bundled_recipe",
        "field_values": {
          "cnn_mean": {
            "model.convolution": "dense",
            "model.coordinates": false,
            "model.ff_ratio": 2,
            "model.heads": "<absent>",
            "model.kernel_size": 3,
            "model.mode": "cnn",
            "model.norm": "layer",
            "model.norm_first": true,
            "model.patch_size": "<absent>",
            "model.pool_grid": 3,
            "model.pooling": "mean",
            "model.position": "<absent>",
            "model.stem_stride": 1,
            "model.width": 24,
            "optimizer.lr": 0.003
          },
          "flat": {
            "model.convolution": "<absent>",
            "model.coordinates": "<absent>",
            "model.ff_ratio": "<absent>",
            "model.heads": "<absent>",
            "model.kernel_size": "<absent>",
            "model.mode": "mlp",
            "model.norm": "<absent>",
            "model.norm_first": "<absent>",
            "model.patch_size": "<absent>",
            "model.pool_grid": "<absent>",
            "model.pooling": "<absent>",
            "model.position": "<absent>",
            "model.stem_stride": "<absent>",
            "model.width": 96,
            "optimizer.lr": 0.001
          },
          "attention": {
            "model.convolution": "<absent>",
            "model.coordinates": "<absent>",
            "model.ff_ratio": 2,
            "model.heads": 4,
            "model.kernel_size": "<absent>",
            "model.mode": "patch_transformer",
            "model.norm": "layer",
            "model.norm_first": true,
            "model.patch_size": 4,
            "model.pool_grid": "<absent>",
            "model.pooling": "<absent>",
            "model.position": "learned",
            "model.stem_stride": "<absent>",
            "model.width": 32,
            "optimizer.lr": 0.001
          }
        },
        "actual_changed_fields": [
          "model.convolution",
          "model.coordinates",
          "model.ff_ratio",
          "model.heads",
          "model.kernel_size",
          "model.mode",
          "model.norm",
          "model.norm_first",
          "model.patch_size",
          "model.pool_grid",
          "model.pooling",
          "model.position",
          "model.stem_stride",
          "model.width",
          "optimizer.lr"
        ]
      },
      "choices": [
        {
          "id": "cnn_mean",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_ec941ffe7c1629588866"
        },
        {
          "id": "flat",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "mlp",
            "width": 96,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 129700,
          "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010",
          "node_id": "candidate_0042c2c253ff4d460405"
        },
        {
          "id": "attention",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_b8f7de70a865c14a6fc0"
        }
      ],
      "diagnostics": [
        {
          "id": "cnn_spatial",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_bfd304b5b5806dae106f"
        }
      ]
    },
    {
      "id": "kernel_stride_texture_s841901",
      "category": "spatial_sampling",
      "context_ref": "textures",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.kernel_size",
        "model.stem_stride"
      ],
      "hypothesis": "Spatial reduction changes texture representation and optimization.",
      "explanation_target": "Use all 2x3 kernel/stride cells. The strided stem is learned patch projection, not a guaranteed anti-aliasing filter.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation.",
        "Additional patch-transformer diagnostic is a preselected learnability anchor, not part of the convolution kernel/stride factorial."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "kernel_stride_texture",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_9f616ad5ea615654",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.kernel_size",
          "model.stem_stride"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "k3_s1": {
            "model.kernel_size": 3,
            "model.stem_stride": 1
          },
          "k5_s2": {
            "model.kernel_size": 5,
            "model.stem_stride": 2
          },
          "k3_s4": {
            "model.kernel_size": 3,
            "model.stem_stride": 4
          }
        },
        "actual_changed_fields": [
          "model.kernel_size",
          "model.stem_stride"
        ]
      },
      "choices": [
        {
          "id": "k3_s1",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_ab616a897d11be58391b"
        },
        {
          "id": "k5_s2",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 5,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 2,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 34012,
          "model_code_sha256": "394a7050f24d49bb1c1fe5a84fddde7fb349d74bb928ebd3d56918f08922ce9e",
          "node_id": "candidate_149005313738d78a4ed7"
        },
        {
          "id": "k3_s4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 4,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16156,
          "model_code_sha256": "247e1e707578507112126325851aaa80fe6e7e885427ce689c887595d7b6e215",
          "node_id": "candidate_aafda8e15a0bbf01406c"
        }
      ],
      "diagnostics": [
        {
          "id": "k5_s1",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 5,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 33868,
          "model_code_sha256": "e9a9a2765f20d57dccca28a3fffdc7dadefe39637b32ed6013d958be3159f651",
          "node_id": "candidate_6e5f0ab5730336704ac8"
        },
        {
          "id": "k3_s2",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 2,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15580,
          "model_code_sha256": "4eed0b51c7ee7ec1f74effe0bdf240b1ba24009b9af627dd4c278268ccdac078",
          "node_id": "candidate_a434611033b8affdc6bd"
        },
        {
          "id": "k5_s4",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 5,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 4,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 34588,
          "model_code_sha256": "df29ae51767411dcd43886969870251fe836bfde85d67f093e5d3a629a91112d",
          "node_id": "candidate_4199c0cb23d1c4c4703a"
        },
        {
          "id": "patch4_learnability_anchor",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_fad13efd0f5a1af0c4eb"
        }
      ]
    },
    {
      "id": "kernel_stride_texture_s841902",
      "category": "spatial_sampling",
      "context_ref": "textures",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.kernel_size",
        "model.stem_stride"
      ],
      "hypothesis": "Spatial reduction changes texture representation and optimization.",
      "explanation_target": "Use all 2x3 kernel/stride cells. The strided stem is learned patch projection, not a guaranteed anti-aliasing filter.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation.",
        "Additional patch-transformer diagnostic is a preselected learnability anchor, not part of the convolution kernel/stride factorial."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "kernel_stride_texture",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_9379541427e4f9a7",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.kernel_size",
          "model.stem_stride"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "k3_s1": {
            "model.kernel_size": 3,
            "model.stem_stride": 1
          },
          "k5_s2": {
            "model.kernel_size": 5,
            "model.stem_stride": 2
          },
          "k3_s4": {
            "model.kernel_size": 3,
            "model.stem_stride": 4
          }
        },
        "actual_changed_fields": [
          "model.kernel_size",
          "model.stem_stride"
        ]
      },
      "choices": [
        {
          "id": "k3_s1",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_37f73a292361df1c6720"
        },
        {
          "id": "k5_s2",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 5,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 2,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 34012,
          "model_code_sha256": "394a7050f24d49bb1c1fe5a84fddde7fb349d74bb928ebd3d56918f08922ce9e",
          "node_id": "candidate_66d46963eeb75fd48934"
        },
        {
          "id": "k3_s4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 4,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16156,
          "model_code_sha256": "247e1e707578507112126325851aaa80fe6e7e885427ce689c887595d7b6e215",
          "node_id": "candidate_f430ce2ad178a479ae16"
        }
      ],
      "diagnostics": [
        {
          "id": "k5_s1",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 5,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 33868,
          "model_code_sha256": "e9a9a2765f20d57dccca28a3fffdc7dadefe39637b32ed6013d958be3159f651",
          "node_id": "candidate_abcdf25ddb5ee63b8869"
        },
        {
          "id": "k3_s2",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 2,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15580,
          "model_code_sha256": "4eed0b51c7ee7ec1f74effe0bdf240b1ba24009b9af627dd4c278268ccdac078",
          "node_id": "candidate_c0578c1bb16eb1b7a0df"
        },
        {
          "id": "k5_s4",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 5,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 4,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 34588,
          "model_code_sha256": "df29ae51767411dcd43886969870251fe836bfde85d67f093e5d3a629a91112d",
          "node_id": "candidate_4934024985276952b5ab"
        },
        {
          "id": "patch4_learnability_anchor",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_6ed7409ffc8cf32c983f"
        }
      ]
    },
    {
      "id": "kernel_stride_texture_s841903",
      "category": "spatial_sampling",
      "context_ref": "textures",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.kernel_size",
        "model.stem_stride"
      ],
      "hypothesis": "Spatial reduction changes texture representation and optimization.",
      "explanation_target": "Use all 2x3 kernel/stride cells. The strided stem is learned patch projection, not a guaranteed anti-aliasing filter.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation.",
        "Additional patch-transformer diagnostic is a preselected learnability anchor, not part of the convolution kernel/stride factorial."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "kernel_stride_texture",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_95967aad9ad199bb",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.kernel_size",
          "model.stem_stride"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "k3_s1": {
            "model.kernel_size": 3,
            "model.stem_stride": 1
          },
          "k5_s2": {
            "model.kernel_size": 5,
            "model.stem_stride": 2
          },
          "k3_s4": {
            "model.kernel_size": 3,
            "model.stem_stride": 4
          }
        },
        "actual_changed_fields": [
          "model.kernel_size",
          "model.stem_stride"
        ]
      },
      "choices": [
        {
          "id": "k3_s1",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_db0952c46eff7da757dd"
        },
        {
          "id": "k5_s2",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 5,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 2,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 34012,
          "model_code_sha256": "394a7050f24d49bb1c1fe5a84fddde7fb349d74bb928ebd3d56918f08922ce9e",
          "node_id": "candidate_7330b39f40c0868f3f45"
        },
        {
          "id": "k3_s4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 4,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16156,
          "model_code_sha256": "247e1e707578507112126325851aaa80fe6e7e885427ce689c887595d7b6e215",
          "node_id": "candidate_b017aaec9762ac9f8f15"
        }
      ],
      "diagnostics": [
        {
          "id": "k5_s1",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 5,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 33868,
          "model_code_sha256": "e9a9a2765f20d57dccca28a3fffdc7dadefe39637b32ed6013d958be3159f651",
          "node_id": "candidate_8e91b1855c6598f64aa6"
        },
        {
          "id": "k3_s2",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 2,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15580,
          "model_code_sha256": "4eed0b51c7ee7ec1f74effe0bdf240b1ba24009b9af627dd4c278268ccdac078",
          "node_id": "candidate_5d1deba868feca1122eb"
        },
        {
          "id": "k5_s4",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 5,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 4,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 34588,
          "model_code_sha256": "df29ae51767411dcd43886969870251fe836bfde85d67f093e5d3a629a91112d",
          "node_id": "candidate_c26909b9e47e00ad719f"
        },
        {
          "id": "patch4_learnability_anchor",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_0bc372c1935b61980671"
        }
      ]
    },
    {
      "id": "patch_high_frequency_s841901",
      "category": "patch",
      "context_ref": "fine_textures",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.patch_size"
      ],
      "hypothesis": "Coarser patches reduce token count but do not inherently erase all within-patch frequency information.",
      "explanation_target": "Patch projection is learned and its output dimension changes compression. Avoid claiming sampling-theorem loss solely from token count.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "patch_high_frequency",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_32df760ef312a38d",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          4.0,
          6.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.patch_size"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "p3": {
            "model.patch_size": 3
          },
          "p4": {
            "model.patch_size": 4
          },
          "p6": {
            "model.patch_size": 6
          }
        },
        "actual_changed_fields": [
          "model.patch_size"
        ]
      },
      "choices": [
        {
          "id": "p3",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 3,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19876,
          "model_code_sha256": "f29b6e6a38c3792bb771f0a25777e143f7555c8c6cf80cd1cd0007277960f5b3",
          "node_id": "candidate_19540c3e091616ae57fb"
        },
        {
          "id": "p4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_3832b1a5bed99bc11e88"
        },
        {
          "id": "p6",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 6,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20068,
          "model_code_sha256": "e71b2de720715f358b0a505a393993f429c24e76a254225b74c00743a9c38227",
          "node_id": "candidate_c36f86452a6409a9d0f0"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "patch_high_frequency_s841902",
      "category": "patch",
      "context_ref": "fine_textures",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.patch_size"
      ],
      "hypothesis": "Coarser patches reduce token count but do not inherently erase all within-patch frequency information.",
      "explanation_target": "Patch projection is learned and its output dimension changes compression. Avoid claiming sampling-theorem loss solely from token count.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "patch_high_frequency",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_e6c4261755f0ff04",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          4.0,
          6.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.patch_size"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "p3": {
            "model.patch_size": 3
          },
          "p4": {
            "model.patch_size": 4
          },
          "p6": {
            "model.patch_size": 6
          }
        },
        "actual_changed_fields": [
          "model.patch_size"
        ]
      },
      "choices": [
        {
          "id": "p3",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 3,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19876,
          "model_code_sha256": "f29b6e6a38c3792bb771f0a25777e143f7555c8c6cf80cd1cd0007277960f5b3",
          "node_id": "candidate_d388f61a539fdadaee0a"
        },
        {
          "id": "p4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_c65741c175d9e7f58b17"
        },
        {
          "id": "p6",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 6,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20068,
          "model_code_sha256": "e71b2de720715f358b0a505a393993f429c24e76a254225b74c00743a9c38227",
          "node_id": "candidate_7b66ca9bca5de7458b8f"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "patch_high_frequency_s841903",
      "category": "patch",
      "context_ref": "fine_textures",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.patch_size"
      ],
      "hypothesis": "Coarser patches reduce token count but do not inherently erase all within-patch frequency information.",
      "explanation_target": "Patch projection is learned and its output dimension changes compression. Avoid claiming sampling-theorem loss solely from token count.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "patch_high_frequency",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_220eb81b568bc599",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          4.0,
          6.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.patch_size"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "p3": {
            "model.patch_size": 3
          },
          "p4": {
            "model.patch_size": 4
          },
          "p6": {
            "model.patch_size": 6
          }
        },
        "actual_changed_fields": [
          "model.patch_size"
        ]
      },
      "choices": [
        {
          "id": "p3",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 3,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19876,
          "model_code_sha256": "f29b6e6a38c3792bb771f0a25777e143f7555c8c6cf80cd1cd0007277960f5b3",
          "node_id": "candidate_ee1139b5d486a75aeb5b"
        },
        {
          "id": "p4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_3afa0c1c9fcc541e8993"
        },
        {
          "id": "p6",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 6,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20068,
          "model_code_sha256": "e71b2de720715f358b0a505a393993f429c24e76a254225b74c00743a9c38227",
          "node_id": "candidate_a0b806c0a1e8a7e9d67a"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "depthwise_width_s841901",
      "category": "block",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.convolution",
        "model.width"
      ],
      "hypothesis": "Cheap spatial mixing can trade capacity for channel width.",
      "explanation_target": "Complete 2x2 convolution/width controls distinguish spatial mixing from width; parameter and FLOP matching are not implied.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "depthwise_width",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_68a832a939c0b770",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.convolution",
          "model.width"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "dense24": {
            "model.convolution": "dense",
            "model.width": 24
          },
          "depthwise24": {
            "model.convolution": "depthwise",
            "model.width": 24
          },
          "depthwise48": {
            "model.convolution": "depthwise",
            "model.width": 48
          }
        },
        "actual_changed_fields": [
          "model.convolution",
          "model.width"
        ]
      },
      "choices": [
        {
          "id": "dense24",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_3b74befeb2c6c9eb61c9"
        },
        {
          "id": "depthwise24",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "depthwise",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 5500,
          "model_code_sha256": "297bdbdefb2459efc98a425849209a52ad5c0084bb3bc416d73fb096109e7e4e",
          "node_id": "candidate_451e3455c06ccd985fa1"
        },
        {
          "id": "depthwise48",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 48,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "depthwise",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20212,
          "model_code_sha256": "c5b04c3c387f9d97858c89cc93dd742f98b4fc635d9ddef274224866d4a35f0a",
          "node_id": "candidate_bcb7d96501c45ac8fc9b"
        }
      ],
      "diagnostics": [
        {
          "id": "dense48",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 48,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 60820,
          "model_code_sha256": "751ce9f6d2824bfbe26fced8a343596c9f2f627a889f39ed9cd667537564058c",
          "node_id": "candidate_ccb91e6df29450d94d92"
        }
      ]
    },
    {
      "id": "depthwise_width_s841902",
      "category": "block",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.convolution",
        "model.width"
      ],
      "hypothesis": "Cheap spatial mixing can trade capacity for channel width.",
      "explanation_target": "Complete 2x2 convolution/width controls distinguish spatial mixing from width; parameter and FLOP matching are not implied.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "depthwise_width",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.convolution",
          "model.width"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "dense24": {
            "model.convolution": "dense",
            "model.width": 24
          },
          "depthwise24": {
            "model.convolution": "depthwise",
            "model.width": 24
          },
          "depthwise48": {
            "model.convolution": "depthwise",
            "model.width": 48
          }
        },
        "actual_changed_fields": [
          "model.convolution",
          "model.width"
        ]
      },
      "choices": [
        {
          "id": "dense24",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_7399cba3f8ca90f400cc"
        },
        {
          "id": "depthwise24",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "depthwise",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 5500,
          "model_code_sha256": "297bdbdefb2459efc98a425849209a52ad5c0084bb3bc416d73fb096109e7e4e",
          "node_id": "candidate_bce264253457a26bed58"
        },
        {
          "id": "depthwise48",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 48,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "depthwise",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20212,
          "model_code_sha256": "c5b04c3c387f9d97858c89cc93dd742f98b4fc635d9ddef274224866d4a35f0a",
          "node_id": "candidate_47be65ef0493e655d027"
        }
      ],
      "diagnostics": [
        {
          "id": "dense48",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 48,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 60820,
          "model_code_sha256": "751ce9f6d2824bfbe26fced8a343596c9f2f627a889f39ed9cd667537564058c",
          "node_id": "candidate_090457311c5d735692bb"
        }
      ]
    },
    {
      "id": "depthwise_width_s841903",
      "category": "block",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.convolution",
        "model.width"
      ],
      "hypothesis": "Cheap spatial mixing can trade capacity for channel width.",
      "explanation_target": "Complete 2x2 convolution/width controls distinguish spatial mixing from width; parameter and FLOP matching are not implied.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "depthwise_width",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.convolution",
          "model.width"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "dense24": {
            "model.convolution": "dense",
            "model.width": 24
          },
          "depthwise24": {
            "model.convolution": "depthwise",
            "model.width": 24
          },
          "depthwise48": {
            "model.convolution": "depthwise",
            "model.width": 48
          }
        },
        "actual_changed_fields": [
          "model.convolution",
          "model.width"
        ]
      },
      "choices": [
        {
          "id": "dense24",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_ec941ffe7c1629588866"
        },
        {
          "id": "depthwise24",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "depthwise",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 5500,
          "model_code_sha256": "297bdbdefb2459efc98a425849209a52ad5c0084bb3bc416d73fb096109e7e4e",
          "node_id": "candidate_a1697d020b7aceb34212"
        },
        {
          "id": "depthwise48",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 48,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "depthwise",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20212,
          "model_code_sha256": "c5b04c3c387f9d97858c89cc93dd742f98b4fc635d9ddef274224866d4a35f0a",
          "node_id": "candidate_54f7bf4b45b4f2cc5dba"
        }
      ],
      "diagnostics": [
        {
          "id": "dense48",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 48,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 60820,
          "model_code_sha256": "751ce9f6d2824bfbe26fced8a343596c9f2f627a889f39ed9cd667537564058c",
          "node_id": "candidate_ebe867dbb00d2d455551"
        }
      ]
    },
    {
      "id": "residual_depth_s841901",
      "category": "residual",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.residual",
        "model.depth"
      ],
      "hypothesis": "Residual paths can change finite-step learnability as depth increases.",
      "explanation_target": "Use all depth/residual cells; improved gradient flow is a hypothesis unless gradient diagnostics establish it.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "residual_depth",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.depth",
          "model.residual"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "plain2": {
            "model.depth": 2,
            "model.residual": false
          },
          "residual6": {
            "model.depth": 6,
            "model.residual": true
          },
          "plain6": {
            "model.depth": 6,
            "model.residual": false
          }
        },
        "actual_changed_fields": [
          "model.depth",
          "model.residual"
        ]
      },
      "choices": [
        {
          "id": "plain2",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "487f99a81970de56cd83e5ed249732dbfa7860e0a5389abfec6823e543153207",
          "node_id": "candidate_5e02b72d5e785e09d855"
        },
        {
          "id": "residual6",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 6,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 46732,
          "model_code_sha256": "7183d61f2bf2188b283e14ce82318d20f5855fcd0a16bc4a57456c2081c89b67",
          "node_id": "candidate_d970952e0279a86f3e85"
        },
        {
          "id": "plain6",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 6,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 46732,
          "model_code_sha256": "7818c319584d6ddd071c6fa738ca50c22811f9f460d2e6e4469a54be52606813",
          "node_id": "candidate_ba0e44517b74430fa307"
        }
      ],
      "diagnostics": [
        {
          "id": "residual2",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_675b0f0945a65c7728d6"
        }
      ]
    },
    {
      "id": "residual_depth_s841902",
      "category": "residual",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.residual",
        "model.depth"
      ],
      "hypothesis": "Residual paths can change finite-step learnability as depth increases.",
      "explanation_target": "Use all depth/residual cells; improved gradient flow is a hypothesis unless gradient diagnostics establish it.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "residual_depth",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.depth",
          "model.residual"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "plain2": {
            "model.depth": 2,
            "model.residual": false
          },
          "residual6": {
            "model.depth": 6,
            "model.residual": true
          },
          "plain6": {
            "model.depth": 6,
            "model.residual": false
          }
        },
        "actual_changed_fields": [
          "model.depth",
          "model.residual"
        ]
      },
      "choices": [
        {
          "id": "plain2",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "487f99a81970de56cd83e5ed249732dbfa7860e0a5389abfec6823e543153207",
          "node_id": "candidate_c9945d9e94130fa6d514"
        },
        {
          "id": "residual6",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 6,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 46732,
          "model_code_sha256": "7183d61f2bf2188b283e14ce82318d20f5855fcd0a16bc4a57456c2081c89b67",
          "node_id": "candidate_9dd05632feeef8396a70"
        },
        {
          "id": "plain6",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 6,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 46732,
          "model_code_sha256": "7818c319584d6ddd071c6fa738ca50c22811f9f460d2e6e4469a54be52606813",
          "node_id": "candidate_3d75613e820affe88db9"
        }
      ],
      "diagnostics": [
        {
          "id": "residual2",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_f9bdfa11ef52bc5e3528"
        }
      ]
    },
    {
      "id": "residual_depth_s841903",
      "category": "residual",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.residual",
        "model.depth"
      ],
      "hypothesis": "Residual paths can change finite-step learnability as depth increases.",
      "explanation_target": "Use all depth/residual cells; improved gradient flow is a hypothesis unless gradient diagnostics establish it.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "residual_depth",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.depth",
          "model.residual"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "plain2": {
            "model.depth": 2,
            "model.residual": false
          },
          "residual6": {
            "model.depth": 6,
            "model.residual": true
          },
          "plain6": {
            "model.depth": 6,
            "model.residual": false
          }
        },
        "actual_changed_fields": [
          "model.depth",
          "model.residual"
        ]
      },
      "choices": [
        {
          "id": "plain2",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "487f99a81970de56cd83e5ed249732dbfa7860e0a5389abfec6823e543153207",
          "node_id": "candidate_ec18f033d414ed8b2267"
        },
        {
          "id": "residual6",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 6,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 46732,
          "model_code_sha256": "7183d61f2bf2188b283e14ce82318d20f5855fcd0a16bc4a57456c2081c89b67",
          "node_id": "candidate_6df66b35fc95baca2b92"
        },
        {
          "id": "plain6",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 6,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 46732,
          "model_code_sha256": "7818c319584d6ddd071c6fa738ca50c22811f9f460d2e6e4469a54be52606813",
          "node_id": "candidate_ce554a6b3b83b800b884"
        }
      ],
      "diagnostics": [
        {
          "id": "residual2",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_d6f536a09652df88c280"
        }
      ]
    },
    {
      "id": "norm_residual_s841901",
      "category": "block",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.norm",
        "model.residual"
      ],
      "hypothesis": "Per-pixel channel normalization and residual paths may interact.",
      "explanation_target": "Complete 2x2 controls; LayerNorm is across channels at each pixel, not BatchNorm or spatial normalization.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "norm_residual",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_68a832a939c0b770",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.norm",
          "model.residual"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "no_norm": {
            "model.norm": "none",
            "model.residual": true
          },
          "plain_norm": {
            "model.norm": "layer",
            "model.residual": false
          },
          "residual_norm": {
            "model.norm": "layer",
            "model.residual": true
          }
        },
        "actual_changed_fields": [
          "model.norm",
          "model.residual"
        ]
      },
      "choices": [
        {
          "id": "no_norm",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "none",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15340,
          "model_code_sha256": "384ea3383ad48c29b344417d8d4ed4a0692433ae1e3d913dceeb3f2764b1ed74",
          "node_id": "candidate_bb0449ecad9dc1b5fc11"
        },
        {
          "id": "plain_norm",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "c08a63248209b0d5a64c9236d54c5be7f1beb598dc26e8c73271614ded80629b",
          "node_id": "candidate_1e525757e242e2130e66"
        },
        {
          "id": "residual_norm",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_3b74befeb2c6c9eb61c9"
        }
      ],
      "diagnostics": [
        {
          "id": "plain_no_norm",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": false,
            "norm": "none",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15340,
          "model_code_sha256": "ea5fa570ff31093c95f665efc19d7c1108483d3113df92e6417e1a5f1fe6a40e",
          "node_id": "candidate_6b94f7549d0848a00d04"
        }
      ]
    },
    {
      "id": "norm_residual_s841902",
      "category": "block",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.norm",
        "model.residual"
      ],
      "hypothesis": "Per-pixel channel normalization and residual paths may interact.",
      "explanation_target": "Complete 2x2 controls; LayerNorm is across channels at each pixel, not BatchNorm or spatial normalization.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "norm_residual",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.norm",
          "model.residual"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "no_norm": {
            "model.norm": "none",
            "model.residual": true
          },
          "plain_norm": {
            "model.norm": "layer",
            "model.residual": false
          },
          "residual_norm": {
            "model.norm": "layer",
            "model.residual": true
          }
        },
        "actual_changed_fields": [
          "model.norm",
          "model.residual"
        ]
      },
      "choices": [
        {
          "id": "no_norm",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "none",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15340,
          "model_code_sha256": "384ea3383ad48c29b344417d8d4ed4a0692433ae1e3d913dceeb3f2764b1ed74",
          "node_id": "candidate_847a0405ced8043e5e4f"
        },
        {
          "id": "plain_norm",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "c08a63248209b0d5a64c9236d54c5be7f1beb598dc26e8c73271614ded80629b",
          "node_id": "candidate_987bbf0bc6a81712d9f8"
        },
        {
          "id": "residual_norm",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_7399cba3f8ca90f400cc"
        }
      ],
      "diagnostics": [
        {
          "id": "plain_no_norm",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": false,
            "norm": "none",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15340,
          "model_code_sha256": "ea5fa570ff31093c95f665efc19d7c1108483d3113df92e6417e1a5f1fe6a40e",
          "node_id": "candidate_9a9a9b9d6908cdc857fc"
        }
      ]
    },
    {
      "id": "norm_residual_s841903",
      "category": "block",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.norm",
        "model.residual"
      ],
      "hypothesis": "Per-pixel channel normalization and residual paths may interact.",
      "explanation_target": "Complete 2x2 controls; LayerNorm is across channels at each pixel, not BatchNorm or spatial normalization.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "norm_residual",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.norm",
          "model.residual"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "no_norm": {
            "model.norm": "none",
            "model.residual": true
          },
          "plain_norm": {
            "model.norm": "layer",
            "model.residual": false
          },
          "residual_norm": {
            "model.norm": "layer",
            "model.residual": true
          }
        },
        "actual_changed_fields": [
          "model.norm",
          "model.residual"
        ]
      },
      "choices": [
        {
          "id": "no_norm",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "none",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15340,
          "model_code_sha256": "384ea3383ad48c29b344417d8d4ed4a0692433ae1e3d913dceeb3f2764b1ed74",
          "node_id": "candidate_fa16bd42cc361c317d1e"
        },
        {
          "id": "plain_norm",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "c08a63248209b0d5a64c9236d54c5be7f1beb598dc26e8c73271614ded80629b",
          "node_id": "candidate_979cc6b964dc1ad05d7a"
        },
        {
          "id": "residual_norm",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_ec941ffe7c1629588866"
        }
      ],
      "diagnostics": [
        {
          "id": "plain_no_norm",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": false,
            "norm": "none",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15340,
          "model_code_sha256": "ea5fa570ff31093c95f665efc19d7c1108483d3113df92e6417e1a5f1fe6a40e",
          "node_id": "candidate_dc7213f3bdc686d92450"
        }
      ]
    },
    {
      "id": "ff_activation_s841901",
      "category": "block",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.ff_ratio",
        "model.activation"
      ],
      "hypothesis": "Feedforward width and activation can trade off in local image blocks.",
      "explanation_target": "Use the full 3x2 width/activation control; width also changes parameter count.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "ff_activation",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_68a832a939c0b770",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.activation",
          "model.ff_ratio"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "ff1_relu": {
            "model.activation": "relu",
            "model.ff_ratio": 1
          },
          "ff2_gelu": {
            "model.activation": "gelu",
            "model.ff_ratio": 2
          },
          "ff4_relu": {
            "model.activation": "relu",
            "model.ff_ratio": 4
          }
        },
        "actual_changed_fields": [
          "model.activation",
          "model.ff_ratio"
        ]
      },
      "choices": [
        {
          "id": "ff1_relu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "relu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 1,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 13084,
          "model_code_sha256": "d9997a6866de1641a44db88c5d3ea82cbc449f26a4ffc4d05ac1a573bbcb0a30",
          "node_id": "candidate_b761d5a572e7c8930b00"
        },
        {
          "id": "ff2_gelu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_3b74befeb2c6c9eb61c9"
        },
        {
          "id": "ff4_relu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "relu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 4,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20140,
          "model_code_sha256": "c3cac7123a9b71d029392ad0bd01da692d128efe5f3f2267300a9d4cd61e5659",
          "node_id": "candidate_99cb23fdd2ff81139333"
        }
      ],
      "diagnostics": [
        {
          "id": "ff1_gelu",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 1,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 13084,
          "model_code_sha256": "1486f09e036eb722ea94783ce64d303a75e37f0d35a5b0010bf6d3362b2bb66c",
          "node_id": "candidate_121eb45f3650881324a7"
        },
        {
          "id": "ff2_relu",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "relu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "a93226785f393655cecb8174ae617a22d94e8251a2ae6e3ea53ad51c9e5b4e75",
          "node_id": "candidate_64fb301b46a587e3fc10"
        },
        {
          "id": "ff4_gelu",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 4,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20140,
          "model_code_sha256": "32cca09dafc2c833b1906f5574e0c99166b7f54576f13dae140418f49f45a2cb",
          "node_id": "candidate_a4320ec04055e927c458"
        }
      ]
    },
    {
      "id": "ff_activation_s841902",
      "category": "block",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.ff_ratio",
        "model.activation"
      ],
      "hypothesis": "Feedforward width and activation can trade off in local image blocks.",
      "explanation_target": "Use the full 3x2 width/activation control; width also changes parameter count.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "ff_activation",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.activation",
          "model.ff_ratio"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "ff1_relu": {
            "model.activation": "relu",
            "model.ff_ratio": 1
          },
          "ff2_gelu": {
            "model.activation": "gelu",
            "model.ff_ratio": 2
          },
          "ff4_relu": {
            "model.activation": "relu",
            "model.ff_ratio": 4
          }
        },
        "actual_changed_fields": [
          "model.activation",
          "model.ff_ratio"
        ]
      },
      "choices": [
        {
          "id": "ff1_relu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "relu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 1,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 13084,
          "model_code_sha256": "d9997a6866de1641a44db88c5d3ea82cbc449f26a4ffc4d05ac1a573bbcb0a30",
          "node_id": "candidate_6b81d3cd60d9b34c45c9"
        },
        {
          "id": "ff2_gelu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_7399cba3f8ca90f400cc"
        },
        {
          "id": "ff4_relu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "relu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 4,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20140,
          "model_code_sha256": "c3cac7123a9b71d029392ad0bd01da692d128efe5f3f2267300a9d4cd61e5659",
          "node_id": "candidate_3692b717bd7aa351ee60"
        }
      ],
      "diagnostics": [
        {
          "id": "ff1_gelu",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 1,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 13084,
          "model_code_sha256": "1486f09e036eb722ea94783ce64d303a75e37f0d35a5b0010bf6d3362b2bb66c",
          "node_id": "candidate_69446f8cd2138a6a629e"
        },
        {
          "id": "ff2_relu",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "relu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "a93226785f393655cecb8174ae617a22d94e8251a2ae6e3ea53ad51c9e5b4e75",
          "node_id": "candidate_614911c0fd261ee640a2"
        },
        {
          "id": "ff4_gelu",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 4,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20140,
          "model_code_sha256": "32cca09dafc2c833b1906f5574e0c99166b7f54576f13dae140418f49f45a2cb",
          "node_id": "candidate_f9a29ada749d4adf5672"
        }
      ]
    },
    {
      "id": "ff_activation_s841903",
      "category": "block",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.ff_ratio",
        "model.activation"
      ],
      "hypothesis": "Feedforward width and activation can trade off in local image blocks.",
      "explanation_target": "Use the full 3x2 width/activation control; width also changes parameter count.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "ff_activation",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.activation",
          "model.ff_ratio"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "ff1_relu": {
            "model.activation": "relu",
            "model.ff_ratio": 1
          },
          "ff2_gelu": {
            "model.activation": "gelu",
            "model.ff_ratio": 2
          },
          "ff4_relu": {
            "model.activation": "relu",
            "model.ff_ratio": 4
          }
        },
        "actual_changed_fields": [
          "model.activation",
          "model.ff_ratio"
        ]
      },
      "choices": [
        {
          "id": "ff1_relu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "relu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 1,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 13084,
          "model_code_sha256": "d9997a6866de1641a44db88c5d3ea82cbc449f26a4ffc4d05ac1a573bbcb0a30",
          "node_id": "candidate_516e799a1c22179b8da9"
        },
        {
          "id": "ff2_gelu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_ec941ffe7c1629588866"
        },
        {
          "id": "ff4_relu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "relu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 4,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20140,
          "model_code_sha256": "c3cac7123a9b71d029392ad0bd01da692d128efe5f3f2267300a9d4cd61e5659",
          "node_id": "candidate_a51f41b8733efb67e1dd"
        }
      ],
      "diagnostics": [
        {
          "id": "ff1_gelu",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 1,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 13084,
          "model_code_sha256": "1486f09e036eb722ea94783ce64d303a75e37f0d35a5b0010bf6d3362b2bb66c",
          "node_id": "candidate_102482d9d6b525ece005"
        },
        {
          "id": "ff2_relu",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "relu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "a93226785f393655cecb8174ae617a22d94e8251a2ae6e3ea53ad51c9e5b4e75",
          "node_id": "candidate_f55c23b4528fe92a3654"
        },
        {
          "id": "ff4_gelu",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 4,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 20140,
          "model_code_sha256": "32cca09dafc2c833b1906f5574e0c99166b7f54576f13dae140418f49f45a2cb",
          "node_id": "candidate_8c794a0d7e554257b434"
        }
      ]
    },
    {
      "id": "attention_heads_ff_s841901",
      "category": "block",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.heads",
        "model.ff_ratio"
      ],
      "hypothesis": "Head count and feedforward allocation can affect finite-budget optimization differently.",
      "explanation_target": "Heads alone keep attention parameter count fixed at fixed width. Controls isolate head and FF effects around the common anchor; not a complete 3x3 grid.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "attention_heads_ff",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.ff_ratio",
          "model.heads"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "h1_ff4": {
            "model.ff_ratio": 4,
            "model.heads": 1
          },
          "h4_ff2": {
            "model.ff_ratio": 2,
            "model.heads": 4
          },
          "h8_ff1": {
            "model.ff_ratio": 1,
            "model.heads": 8
          }
        },
        "actual_changed_fields": [
          "model.ff_ratio",
          "model.heads"
        ]
      },
      "choices": [
        {
          "id": "h1_ff4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 4,
            "heads": 1,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 27748,
          "model_code_sha256": "33aa4418bafc50c1f23adb582ed959f04f2cfe8daf97db6cd775f669564215b7",
          "node_id": "candidate_d315cbf6c5d694184220"
        },
        {
          "id": "h4_ff2",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_5cf50db65aa67995a4f9"
        },
        {
          "id": "h8_ff1",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 1,
            "heads": 8,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15268,
          "model_code_sha256": "37b0b123040119e40aa10bdfce698d951381255500bad19d4d734cb820a3bc1d",
          "node_id": "candidate_d0e587aed7b3c0b52997"
        }
      ],
      "diagnostics": [
        {
          "id": "h1_ff2",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 1,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "4038255caa48fa6b4fa7829427ab55fcf105ad98de0ac407c33ee34def5a043a",
          "node_id": "candidate_f1ea60b4a966c158b63c"
        },
        {
          "id": "h8_ff2",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 8,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "9e76b05e6a7a0f7d75de4c9fd9ac4652a7ce37e4f12081897317f660e7ca7621",
          "node_id": "candidate_8630080f46d5bcc4c4fa"
        },
        {
          "id": "h4_ff1",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 1,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15268,
          "model_code_sha256": "8364bbae844edba18d6067aa23e9bf7faaded3ed3722211383bbd3e63c55c6da",
          "node_id": "candidate_699c6c5c2d2eb10b3040"
        },
        {
          "id": "h4_ff4",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 4,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 27748,
          "model_code_sha256": "5e4a2c32ed8426dea1fabecda2b813b8ff8e4a7d0fcafbd8be819520e625e953",
          "node_id": "candidate_cdfd16b183904fe63c5a"
        }
      ]
    },
    {
      "id": "attention_heads_ff_s841902",
      "category": "block",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.heads",
        "model.ff_ratio"
      ],
      "hypothesis": "Head count and feedforward allocation can affect finite-budget optimization differently.",
      "explanation_target": "Heads alone keep attention parameter count fixed at fixed width. Controls isolate head and FF effects around the common anchor; not a complete 3x3 grid.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "attention_heads_ff",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.ff_ratio",
          "model.heads"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "h1_ff4": {
            "model.ff_ratio": 4,
            "model.heads": 1
          },
          "h4_ff2": {
            "model.ff_ratio": 2,
            "model.heads": 4
          },
          "h8_ff1": {
            "model.ff_ratio": 1,
            "model.heads": 8
          }
        },
        "actual_changed_fields": [
          "model.ff_ratio",
          "model.heads"
        ]
      },
      "choices": [
        {
          "id": "h1_ff4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 4,
            "heads": 1,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 27748,
          "model_code_sha256": "33aa4418bafc50c1f23adb582ed959f04f2cfe8daf97db6cd775f669564215b7",
          "node_id": "candidate_2e250ec49da02dabf3e9"
        },
        {
          "id": "h4_ff2",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_5313f05220195b0f67e9"
        },
        {
          "id": "h8_ff1",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 1,
            "heads": 8,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15268,
          "model_code_sha256": "37b0b123040119e40aa10bdfce698d951381255500bad19d4d734cb820a3bc1d",
          "node_id": "candidate_09a8640fa4d7c593644f"
        }
      ],
      "diagnostics": [
        {
          "id": "h1_ff2",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 1,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "4038255caa48fa6b4fa7829427ab55fcf105ad98de0ac407c33ee34def5a043a",
          "node_id": "candidate_2d28083738832680f9be"
        },
        {
          "id": "h8_ff2",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 8,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "9e76b05e6a7a0f7d75de4c9fd9ac4652a7ce37e4f12081897317f660e7ca7621",
          "node_id": "candidate_2e24746f420e07884c4c"
        },
        {
          "id": "h4_ff1",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 1,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15268,
          "model_code_sha256": "8364bbae844edba18d6067aa23e9bf7faaded3ed3722211383bbd3e63c55c6da",
          "node_id": "candidate_a301024fb64cd6869bd2"
        },
        {
          "id": "h4_ff4",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 4,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 27748,
          "model_code_sha256": "5e4a2c32ed8426dea1fabecda2b813b8ff8e4a7d0fcafbd8be819520e625e953",
          "node_id": "candidate_db3bac963e47c6104d6a"
        }
      ]
    },
    {
      "id": "attention_heads_ff_s841903",
      "category": "block",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.heads",
        "model.ff_ratio"
      ],
      "hypothesis": "Head count and feedforward allocation can affect finite-budget optimization differently.",
      "explanation_target": "Heads alone keep attention parameter count fixed at fixed width. Controls isolate head and FF effects around the common anchor; not a complete 3x3 grid.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "attention_heads_ff",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.ff_ratio",
          "model.heads"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "h1_ff4": {
            "model.ff_ratio": 4,
            "model.heads": 1
          },
          "h4_ff2": {
            "model.ff_ratio": 2,
            "model.heads": 4
          },
          "h8_ff1": {
            "model.ff_ratio": 1,
            "model.heads": 8
          }
        },
        "actual_changed_fields": [
          "model.ff_ratio",
          "model.heads"
        ]
      },
      "choices": [
        {
          "id": "h1_ff4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 4,
            "heads": 1,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 27748,
          "model_code_sha256": "33aa4418bafc50c1f23adb582ed959f04f2cfe8daf97db6cd775f669564215b7",
          "node_id": "candidate_c205e04736f94326026a"
        },
        {
          "id": "h4_ff2",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_681cbfb92210830d5199"
        },
        {
          "id": "h8_ff1",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 1,
            "heads": 8,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15268,
          "model_code_sha256": "37b0b123040119e40aa10bdfce698d951381255500bad19d4d734cb820a3bc1d",
          "node_id": "candidate_b42338d48a1242eb97d9"
        }
      ],
      "diagnostics": [
        {
          "id": "h1_ff2",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 1,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "4038255caa48fa6b4fa7829427ab55fcf105ad98de0ac407c33ee34def5a043a",
          "node_id": "candidate_58d373fe55636c9acd59"
        },
        {
          "id": "h8_ff2",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "heads": 8,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "9e76b05e6a7a0f7d75de4c9fd9ac4652a7ce37e4f12081897317f660e7ca7621",
          "node_id": "candidate_229ab15aa9cbfc005fb3"
        },
        {
          "id": "h4_ff1",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 1,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15268,
          "model_code_sha256": "8364bbae844edba18d6067aa23e9bf7faaded3ed3722211383bbd3e63c55c6da",
          "node_id": "candidate_c0eff0c920f37eab36d3"
        },
        {
          "id": "h4_ff4",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "patch_transformer",
            "width": 32,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 4,
            "heads": 4,
            "patch_size": 4,
            "position": "learned",
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 27748,
          "model_code_sha256": "5e4a2c32ed8426dea1fabecda2b813b8ff8e4a7d0fcafbd8be819520e625e953",
          "node_id": "candidate_515b18e0245f3d9b63bd"
        }
      ]
    },
    {
      "id": "learning_rate_s841901",
      "category": "training",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "optimizer.lr"
      ],
      "hypothesis": "A fixed training budget can expose both undertraining and unstable optimization.",
      "explanation_target": "Use actual train/held-out trajectories; a high learning rate is not itself proof of divergence.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "learning_rate",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_68a832a939c0b770",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "optimizer.lr"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "lr1e4": {
            "optimizer.lr": 0.0001
          },
          "lr1e3": {
            "optimizer.lr": 0.001
          },
          "lr1e2": {
            "optimizer.lr": 0.01
          }
        },
        "actual_changed_fields": [
          "optimizer.lr"
        ]
      },
      "choices": [
        {
          "id": "lr1e4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.0001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_31abb7f5ae858e9f35e8"
        },
        {
          "id": "lr1e3",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_6ec59d4186f9b42d6754"
        },
        {
          "id": "lr1e2",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.01,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_b07779969f06e2a919a4"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "learning_rate_s841902",
      "category": "training",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "optimizer.lr"
      ],
      "hypothesis": "A fixed training budget can expose both undertraining and unstable optimization.",
      "explanation_target": "Use actual train/held-out trajectories; a high learning rate is not itself proof of divergence.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "learning_rate",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_acbb3cacc35c6e86",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "optimizer.lr"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "lr1e4": {
            "optimizer.lr": 0.0001
          },
          "lr1e3": {
            "optimizer.lr": 0.001
          },
          "lr1e2": {
            "optimizer.lr": 0.01
          }
        },
        "actual_changed_fields": [
          "optimizer.lr"
        ]
      },
      "choices": [
        {
          "id": "lr1e4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.0001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_68c3418242c7dd2f2c5e"
        },
        {
          "id": "lr1e3",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_3e9dff76e192bdb82479"
        },
        {
          "id": "lr1e2",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.01,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_878848a639e130ad11ad"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "learning_rate_s841903",
      "category": "training",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "optimizer.lr"
      ],
      "hypothesis": "A fixed training budget can expose both undertraining and unstable optimization.",
      "explanation_target": "Use actual train/held-out trajectories; a high learning rate is not itself proof of divergence.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "learning_rate",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_43d6af5f6d3b729d",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "optimizer.lr"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "lr1e4": {
            "optimizer.lr": 0.0001
          },
          "lr1e3": {
            "optimizer.lr": 0.001
          },
          "lr1e2": {
            "optimizer.lr": 0.01
          }
        },
        "actual_changed_fields": [
          "optimizer.lr"
        ]
      },
      "choices": [
        {
          "id": "lr1e4",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.0001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_ab7b2bbaecc26528fc3a"
        },
        {
          "id": "lr1e3",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_e313c07c685170294dbd"
        },
        {
          "id": "lr1e2",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.01,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_d8544172036bc9ae086b"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "decay_small_data_s841901",
      "category": "training",
      "context_ref": "small_shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "optimizer.weight_decay"
      ],
      "hypothesis": "Weight decay interacts with a small clean training set and finite steps.",
      "explanation_target": "Distinguish underfitting from generalization with training and held-out losses; no label-noise explanation is admissible because labels are exact.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "decay_small_data",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_4d28614559aee4b9",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 64,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "optimizer.weight_decay"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "wd0": {
            "optimizer.weight_decay": 0.0
          },
          "wd03": {
            "optimizer.weight_decay": 0.03
          },
          "wd3": {
            "optimizer.weight_decay": 0.3
          }
        },
        "actual_changed_fields": [
          "optimizer.weight_decay"
        ]
      },
      "choices": [
        {
          "id": "wd0",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.0,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_bea9b669d36ce0bb0f49"
        },
        {
          "id": "wd03",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.03,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_bf1af3070b43afb13889"
        },
        {
          "id": "wd3",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.3,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_69c2b002b73676a9e693"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "decay_small_data_s841902",
      "category": "training",
      "context_ref": "small_shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "optimizer.weight_decay"
      ],
      "hypothesis": "Weight decay interacts with a small clean training set and finite steps.",
      "explanation_target": "Distinguish underfitting from generalization with training and held-out losses; no label-noise explanation is admissible because labels are exact.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "decay_small_data",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_484bdc518daa14bb",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 64,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "optimizer.weight_decay"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "wd0": {
            "optimizer.weight_decay": 0.0
          },
          "wd03": {
            "optimizer.weight_decay": 0.03
          },
          "wd3": {
            "optimizer.weight_decay": 0.3
          }
        },
        "actual_changed_fields": [
          "optimizer.weight_decay"
        ]
      },
      "choices": [
        {
          "id": "wd0",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.0,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_bbcaf6772f71b9bfe1cc"
        },
        {
          "id": "wd03",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.03,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_decaf5a0597804014f47"
        },
        {
          "id": "wd3",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.3,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_5449f78632b714ebdda4"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "decay_small_data_s841903",
      "category": "training",
      "context_ref": "small_shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "optimizer.weight_decay"
      ],
      "hypothesis": "Weight decay interacts with a small clean training set and finite steps.",
      "explanation_target": "Distinguish underfitting from generalization with training and held-out losses; no label-noise explanation is admissible because labels are exact.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "decay_small_data",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_63f24d2a7f7dd6e2",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 64,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "optimizer.weight_decay"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "wd0": {
            "optimizer.weight_decay": 0.0
          },
          "wd03": {
            "optimizer.weight_decay": 0.03
          },
          "wd3": {
            "optimizer.weight_decay": 0.3
          }
        },
        "actual_changed_fields": [
          "optimizer.weight_decay"
        ]
      },
      "choices": [
        {
          "id": "wd0",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.0,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_b998d6e03b93bb19946a"
        },
        {
          "id": "wd03",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.03,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_a610f761f13d94f88c47"
        },
        {
          "id": "wd3",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.3,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_13fb3cf7b2dd483e7ad7"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "batch_fixed_samples_s841901",
      "category": "training",
      "context_ref": "relations",
      "budget_semantics": "matched_samples",
      "varying_axes": [
        "training.batch_size",
        "training.training_steps"
      ],
      "hypothesis": "Equal examples processed do not imply equal update counts or gradient noise.",
      "explanation_target": "Steps are 600/300/150 and each choice processes 9600 examples. This is a coupled batch/update intervention; paired streams do not make different batch sizes identical trajectories.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "batch_fixed_samples",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_samples",
        "changed_fields": [
          "training.batch_size",
          "training.training_steps"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "batch16": {
            "training.batch_size": 16,
            "training.training_steps": 600
          },
          "batch32": {
            "training.batch_size": 32,
            "training.training_steps": 300
          },
          "batch64": {
            "training.batch_size": 64,
            "training.training_steps": 150
          }
        },
        "actual_changed_fields": [
          "training.batch_size",
          "training.training_steps"
        ]
      },
      "choices": [
        {
          "id": "batch16",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 16,
            "training_steps": 600,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              150,
              300,
              600
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_cc8a9186f70e19a657b1"
        },
        {
          "id": "batch32",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_675b0f0945a65c7728d6"
        },
        {
          "id": "batch64",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 64,
            "training_steps": 150,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              37,
              75,
              150
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_f59241db8842e4e80120"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "batch_fixed_samples_s841902",
      "category": "training",
      "context_ref": "relations",
      "budget_semantics": "matched_samples",
      "varying_axes": [
        "training.batch_size",
        "training.training_steps"
      ],
      "hypothesis": "Equal examples processed do not imply equal update counts or gradient noise.",
      "explanation_target": "Steps are 600/300/150 and each choice processes 9600 examples. This is a coupled batch/update intervention; paired streams do not make different batch sizes identical trajectories.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "batch_fixed_samples",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_samples",
        "changed_fields": [
          "training.batch_size",
          "training.training_steps"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "batch16": {
            "training.batch_size": 16,
            "training.training_steps": 600
          },
          "batch32": {
            "training.batch_size": 32,
            "training.training_steps": 300
          },
          "batch64": {
            "training.batch_size": 64,
            "training.training_steps": 150
          }
        },
        "actual_changed_fields": [
          "training.batch_size",
          "training.training_steps"
        ]
      },
      "choices": [
        {
          "id": "batch16",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 16,
            "training_steps": 600,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              150,
              300,
              600
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_7a695f3edd302f57ca0e"
        },
        {
          "id": "batch32",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_f9bdfa11ef52bc5e3528"
        },
        {
          "id": "batch64",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 64,
            "training_steps": 150,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              37,
              75,
              150
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_c54739bfbb790f1d5694"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "batch_fixed_samples_s841903",
      "category": "training",
      "context_ref": "relations",
      "budget_semantics": "matched_samples",
      "varying_axes": [
        "training.batch_size",
        "training.training_steps"
      ],
      "hypothesis": "Equal examples processed do not imply equal update counts or gradient noise.",
      "explanation_target": "Steps are 600/300/150 and each choice processes 9600 examples. This is a coupled batch/update intervention; paired streams do not make different batch sizes identical trajectories.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "batch_fixed_samples",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_samples",
        "changed_fields": [
          "training.batch_size",
          "training.training_steps"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "batch16": {
            "training.batch_size": 16,
            "training.training_steps": 600
          },
          "batch32": {
            "training.batch_size": 32,
            "training.training_steps": 300
          },
          "batch64": {
            "training.batch_size": 64,
            "training.training_steps": 150
          }
        },
        "actual_changed_fields": [
          "training.batch_size",
          "training.training_steps"
        ]
      },
      "choices": [
        {
          "id": "batch16",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 16,
            "training_steps": 600,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              150,
              300,
              600
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_b08890c32275d0926859"
        },
        {
          "id": "batch32",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 300,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              75,
              150,
              300
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_d6f536a09652df88c280"
        },
        {
          "id": "batch64",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 64,
            "training_steps": 150,
            "total_samples_seen": 9600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              37,
              75,
              150
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_df03dcb5bf390a018ea3"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "budget_small_data_s841901",
      "category": "training_budget",
      "context_ref": "small_shapes",
      "budget_semantics": "per_choice_steps",
      "varying_axes": [
        "training.training_steps"
      ],
      "hypothesis": "More updates can improve fitting while worsening held-out calibration on small clean datasets.",
      "explanation_target": "Use exact shared initial states and training prefixes. Report classification error as well as CE; overconfidence can change CE without changing accuracy.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "budget_small_data",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_4d28614559aee4b9",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 64,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "per_choice_steps",
        "changed_fields": [
          "training.training_steps"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "steps100": {
            "training.training_steps": 100
          },
          "steps400": {
            "training.training_steps": 400
          },
          "steps1200": {
            "training.training_steps": 1200
          }
        },
        "actual_changed_fields": [
          "training.training_steps"
        ]
      },
      "choices": [
        {
          "id": "steps100",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 100,
            "total_samples_seen": 3200,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              25,
              50,
              100
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_d6d8d7db50c27549e210"
        },
        {
          "id": "steps400",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 400,
            "total_samples_seen": 12800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              100,
              200,
              400
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_87ff1c77bb033dc6bcf6"
        },
        {
          "id": "steps1200",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_9c4ac004a68ae6f01a23"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "budget_small_data_s841902",
      "category": "training_budget",
      "context_ref": "small_shapes",
      "budget_semantics": "per_choice_steps",
      "varying_axes": [
        "training.training_steps"
      ],
      "hypothesis": "More updates can improve fitting while worsening held-out calibration on small clean datasets.",
      "explanation_target": "Use exact shared initial states and training prefixes. Report classification error as well as CE; overconfidence can change CE without changing accuracy.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "budget_small_data",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_484bdc518daa14bb",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 64,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "per_choice_steps",
        "changed_fields": [
          "training.training_steps"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "steps100": {
            "training.training_steps": 100
          },
          "steps400": {
            "training.training_steps": 400
          },
          "steps1200": {
            "training.training_steps": 1200
          }
        },
        "actual_changed_fields": [
          "training.training_steps"
        ]
      },
      "choices": [
        {
          "id": "steps100",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 100,
            "total_samples_seen": 3200,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              25,
              50,
              100
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_f4212fdb6215bc33b10a"
        },
        {
          "id": "steps400",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 400,
            "total_samples_seen": 12800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              100,
              200,
              400
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_1c8f2761b5434f3d3823"
        },
        {
          "id": "steps1200",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_24ff1215b04f7b60b52c"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "budget_small_data_s841903",
      "category": "training_budget",
      "context_ref": "small_shapes",
      "budget_semantics": "per_choice_steps",
      "varying_axes": [
        "training.training_steps"
      ],
      "hypothesis": "More updates can improve fitting while worsening held-out calibration on small clean datasets.",
      "explanation_target": "Use exact shared initial states and training prefixes. Report classification error as well as CE; overconfidence can change CE without changing accuracy.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "budget_small_data",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_63f24d2a7f7dd6e2",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "local_shape",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 64,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "per_choice_steps",
        "changed_fields": [
          "training.training_steps"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "steps100": {
            "training.training_steps": 100
          },
          "steps400": {
            "training.training_steps": 400
          },
          "steps1200": {
            "training.training_steps": 1200
          }
        },
        "actual_changed_fields": [
          "training.training_steps"
        ]
      },
      "choices": [
        {
          "id": "steps100",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 100,
            "total_samples_seen": 3200,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              25,
              50,
              100
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_004d93eda1cef7d7dbf7"
        },
        {
          "id": "steps400",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 400,
            "total_samples_seen": 12800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              100,
              200,
              400
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_fe3235aaee5ac5aac819"
        },
        {
          "id": "steps1200",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_c33cc593b59cf479dc82"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "kernel_budget_s841901",
      "category": "spatial_budget",
      "context_ref": "relations",
      "budget_semantics": "per_choice_steps",
      "varying_axes": [
        "model.kernel_size",
        "training.training_steps"
      ],
      "hypothesis": "Spatial support and update budget can produce misleading architecture-only narratives.",
      "explanation_target": "Use full 2x3 kernel/budget controls; compare at matched steps before attributing a joint recipe win to receptive field.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "kernel_budget",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_e0bb4b3a905a5977",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "per_choice_steps",
        "changed_fields": [
          "model.kernel_size",
          "training.training_steps"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "k3_short": {
            "model.kernel_size": 3,
            "training.training_steps": 100
          },
          "k7_medium": {
            "model.kernel_size": 7,
            "training.training_steps": 400
          },
          "k3_long": {
            "model.kernel_size": 3,
            "training.training_steps": 800
          }
        },
        "actual_changed_fields": [
          "model.kernel_size",
          "training.training_steps"
        ]
      },
      "choices": [
        {
          "id": "k3_short",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 100,
            "total_samples_seen": 3200,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              25,
              50,
              100
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_1d3aa2831c81134e77af"
        },
        {
          "id": "k7_medium",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 400,
            "total_samples_seen": 12800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              100,
              200,
              400
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 62284,
          "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91",
          "node_id": "candidate_d8494e4bbb475abc8d6c"
        },
        {
          "id": "k3_long",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 800,
            "total_samples_seen": 25600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              200,
              400,
              800
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_5ce646d0e89d3b097492"
        }
      ],
      "diagnostics": [
        {
          "id": "k7_short",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 100,
            "total_samples_seen": 3200,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              25,
              50,
              100
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 62284,
          "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91",
          "node_id": "candidate_fffc5bdace486a294aea"
        },
        {
          "id": "k3_medium",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 400,
            "total_samples_seen": 12800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              100,
              200,
              400
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_2b14bf0c3a6914ae7431"
        },
        {
          "id": "k7_long",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 800,
            "total_samples_seen": 25600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              200,
              400,
              800
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 62284,
          "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91",
          "node_id": "candidate_8ad93f31861df974db61"
        }
      ]
    },
    {
      "id": "kernel_budget_s841902",
      "category": "spatial_budget",
      "context_ref": "relations",
      "budget_semantics": "per_choice_steps",
      "varying_axes": [
        "model.kernel_size",
        "training.training_steps"
      ],
      "hypothesis": "Spatial support and update budget can produce misleading architecture-only narratives.",
      "explanation_target": "Use full 2x3 kernel/budget controls; compare at matched steps before attributing a joint recipe win to receptive field.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "kernel_budget",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_f7fa567d4761b40c",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "per_choice_steps",
        "changed_fields": [
          "model.kernel_size",
          "training.training_steps"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "k3_short": {
            "model.kernel_size": 3,
            "training.training_steps": 100
          },
          "k7_medium": {
            "model.kernel_size": 7,
            "training.training_steps": 400
          },
          "k3_long": {
            "model.kernel_size": 3,
            "training.training_steps": 800
          }
        },
        "actual_changed_fields": [
          "model.kernel_size",
          "training.training_steps"
        ]
      },
      "choices": [
        {
          "id": "k3_short",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 100,
            "total_samples_seen": 3200,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              25,
              50,
              100
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_1ee1c85b02c12c69049c"
        },
        {
          "id": "k7_medium",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 400,
            "total_samples_seen": 12800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              100,
              200,
              400
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 62284,
          "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91",
          "node_id": "candidate_fa921f9d94e2ee00d414"
        },
        {
          "id": "k3_long",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 800,
            "total_samples_seen": 25600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              200,
              400,
              800
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_5bcc6eb2db18941f07c1"
        }
      ],
      "diagnostics": [
        {
          "id": "k7_short",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 100,
            "total_samples_seen": 3200,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              25,
              50,
              100
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 62284,
          "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91",
          "node_id": "candidate_a68f85d27604a60fcfac"
        },
        {
          "id": "k3_medium",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 400,
            "total_samples_seen": 12800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              100,
              200,
              400
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_a9c85c0bf30c5c45b895"
        },
        {
          "id": "k7_long",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 800,
            "total_samples_seen": 25600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              200,
              400,
              800
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 62284,
          "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91",
          "node_id": "candidate_b0777a15e3036a2b0264"
        }
      ]
    },
    {
      "id": "kernel_budget_s841903",
      "category": "spatial_budget",
      "context_ref": "relations",
      "budget_semantics": "per_choice_steps",
      "varying_axes": [
        "model.kernel_size",
        "training.training_steps"
      ],
      "hypothesis": "Spatial support and update budget can produce misleading architecture-only narratives.",
      "explanation_target": "Use full 2x3 kernel/budget controls; compare at matched steps before attributing a joint recipe win to receptive field.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "kernel_budget",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_1544b3cbd88f855f",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "relative_position",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          1.5,
          3.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "per_choice_steps",
        "changed_fields": [
          "model.kernel_size",
          "training.training_steps"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "k3_short": {
            "model.kernel_size": 3,
            "training.training_steps": 100
          },
          "k7_medium": {
            "model.kernel_size": 7,
            "training.training_steps": 400
          },
          "k3_long": {
            "model.kernel_size": 3,
            "training.training_steps": 800
          }
        },
        "actual_changed_fields": [
          "model.kernel_size",
          "training.training_steps"
        ]
      },
      "choices": [
        {
          "id": "k3_short",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 100,
            "total_samples_seen": 3200,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              25,
              50,
              100
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_44d23e55659364508d97"
        },
        {
          "id": "k7_medium",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 400,
            "total_samples_seen": 12800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              100,
              200,
              400
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 62284,
          "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91",
          "node_id": "candidate_dad86d2f0763eb0794d8"
        },
        {
          "id": "k3_long",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 800,
            "total_samples_seen": 25600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              200,
              400,
              800
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_80324d542043d4a3c3d1"
        }
      ],
      "diagnostics": [
        {
          "id": "k7_short",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 100,
            "total_samples_seen": 3200,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              25,
              50,
              100
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 62284,
          "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91",
          "node_id": "candidate_72b2c7b6e515710cb184"
        },
        {
          "id": "k3_medium",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 400,
            "total_samples_seen": 12800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              100,
              200,
              400
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_70fd6488ca27a88be005"
        },
        {
          "id": "k7_long",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 7,
            "convolution": "dense",
            "pooling": "spatial",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.001,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 800,
            "total_samples_seen": 25600,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              200,
              400,
              800
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 62284,
          "model_code_sha256": "677118c835caa5fb30fec234d0d485ba5e9c0a91d9ee488f699205408ca86f91",
          "node_id": "candidate_903f34d3ec924e3bf6a9"
        }
      ]
    },
    {
      "id": "activation_texture_s841901",
      "category": "activation",
      "context_ref": "fine_textures",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.activation"
      ],
      "hypothesis": "Activation affects optimization of a high-frequency visual signal.",
      "explanation_target": "All labels are exact orientation classes. Saturation is a mechanism hypothesis; use activation/gradient measurements before making a unique causal claim.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "activation_texture",
      "instance_seed": 841901,
      "dataset_node_id": "dataset_32df760ef312a38d",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          4.0,
          6.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.activation"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "relu": {
            "model.activation": "relu"
          },
          "gelu": {
            "model.activation": "gelu"
          },
          "tanh": {
            "model.activation": "tanh"
          }
        },
        "actual_changed_fields": [
          "model.activation"
        ]
      },
      "choices": [
        {
          "id": "relu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "relu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "a93226785f393655cecb8174ae617a22d94e8251a2ae6e3ea53ad51c9e5b4e75",
          "node_id": "candidate_159df3b6e29ac46bd076"
        },
        {
          "id": "gelu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_bc1c99efb458f1abecac"
        },
        {
          "id": "tanh",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "tanh",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "101fba5c57246c0fc592fa5b1a30cf7fd569ba0a8001f23a37531234618c32a8",
          "node_id": "candidate_f86a78da04bbe1e05364"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "activation_texture_s841902",
      "category": "activation",
      "context_ref": "fine_textures",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.activation"
      ],
      "hypothesis": "Activation affects optimization of a high-frequency visual signal.",
      "explanation_target": "All labels are exact orientation classes. Saturation is a mechanism hypothesis; use activation/gradient measurements before making a unique causal claim.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "activation_texture",
      "instance_seed": 841902,
      "dataset_node_id": "dataset_e6c4261755f0ff04",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          4.0,
          6.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.activation"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "relu": {
            "model.activation": "relu"
          },
          "gelu": {
            "model.activation": "gelu"
          },
          "tanh": {
            "model.activation": "tanh"
          }
        },
        "actual_changed_fields": [
          "model.activation"
        ]
      },
      "choices": [
        {
          "id": "relu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "relu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "a93226785f393655cecb8174ae617a22d94e8251a2ae6e3ea53ad51c9e5b4e75",
          "node_id": "candidate_ed8b5935d7cedac78147"
        },
        {
          "id": "gelu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_e23d523074ffbb679c5c"
        },
        {
          "id": "tanh",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "tanh",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "101fba5c57246c0fc592fa5b1a30cf7fd569ba0a8001f23a37531234618c32a8",
          "node_id": "candidate_680d4f9ac1b9852312b9"
        }
      ],
      "diagnostics": []
    },
    {
      "id": "activation_texture_s841903",
      "category": "activation",
      "context_ref": "fine_textures",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.activation"
      ],
      "hypothesis": "Activation affects optimization of a high-frequency visual signal.",
      "explanation_target": "All labels are exact orientation classes. Saturation is a mechanism hypothesis; use activation/gradient measurements before making a unique causal claim.",
      "controls": [
        "Same fixed materialized image splits within each question.",
        "All unlisted fields inherit the explicit defaults.",
        "Measured parameter counts are displayed; matching width does not match parameters or computation."
      ],
      "shortcut_checks": [
        "larger model always wins",
        "more steps always wins",
        "matching task name to architecture"
      ],
      "admission": "Freeze discovery-ranked primary triple before independent confirmation. Never replace choices after rejection.",
      "template_id": "activation_texture",
      "instance_seed": 841903,
      "dataset_node_id": "dataset_220eb81b568bc599",
      "dataset_config": {
        "size": 24,
        "num_classes": 4,
        "task": "texture_orientation",
        "radius_range": [
          2.5,
          4.5
        ],
        "frequency_range": [
          4.0,
          6.0
        ],
        "intensity_range": [
          0.8,
          1.2
        ],
        "normalize_mass": true,
        "train_size": 1024,
        "validation_size": 512,
        "test_size": 512,
        "ood_size": 128,
        "edge_softness": 0.5
      },
      "comparison": {
        "budget_semantics": "matched_steps",
        "changed_fields": [
          "model.activation"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "relu": {
            "model.activation": "relu"
          },
          "gelu": {
            "model.activation": "gelu"
          },
          "tanh": {
            "model.activation": "tanh"
          }
        },
        "actual_changed_fields": [
          "model.activation"
        ]
      },
      "choices": [
        {
          "id": "relu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "relu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "a93226785f393655cecb8174ae617a22d94e8251a2ae6e3ea53ad51c9e5b4e75",
          "node_id": "candidate_f685aa68d7fa486641ee"
        },
        {
          "id": "gelu",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "573f9c6f91dc592a90d580cdd9ebf4ef4f5c43edd97b9c1bb2c1580512c42d9f",
          "node_id": "candidate_f6a336f7bc57e0c58080"
        },
        {
          "id": "tanh",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "tanh",
            "residual": true,
            "norm": "layer",
            "norm_first": true,
            "ff_ratio": 2,
            "kernel_size": 3,
            "convolution": "dense",
            "pooling": "mean",
            "pool_grid": 3,
            "coordinates": false,
            "stem_stride": 1,
            "max_parameters": 2000000,
            "input_shape": [
              2,
              24,
              24
            ],
            "output_shape": [
              4
            ],
            "output_kind": "classification"
          },
          "optimizer": {
            "type": "AdamW",
            "lr": 0.003,
            "weight_decay": 0.01,
            "betas": [
              0.9,
              0.999
            ],
            "eps": 1e-08,
            "foreach": false
          },
          "loss": {
            "loss_id": "cross_entropy"
          },
          "training": {
            "batch_size": 32,
            "training_steps": 1200,
            "total_samples_seen": 38400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              300,
              600,
              1200
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 15436,
          "model_code_sha256": "101fba5c57246c0fc592fa5b1a30cf7fd569ba0a8001f23a37531234618c32a8",
          "node_id": "candidate_56051496b194b14c493a"
        }
      ],
      "diagnostics": []
    }
  ]
}