{
  "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-supplement-v1.yaml",
  "source_sha256": "d413b037bf7f9e8f8ab07d25400a54da333593b3be728c6ad07e6c368a1c1b30",
  "base_profile": "profiles/task-expansion-image-classification.yaml",
  "default_num_choices": 3,
  "discovery_training_seeds": [
    300,
    301,
    302,
    303
  ],
  "confirmation_training_seeds": [
    400,
    401,
    402,
    403,
    404,
    405,
    406,
    407,
    408,
    409,
    410,
    411,
    412,
    413,
    414,
    415
  ],
  "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 exactly 24 frozen adjacent hypotheses from all 12 primary triples in this independent supplement, including rejected instances.",
    "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 every rejected triple without replacing choices, expanding templates or tuning against model answers."
  },
  "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": [
        300,
        301,
        302,
        303
      ],
      "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": [
        400,
        401,
        402,
        403,
        404,
        405,
        406,
        407,
        408,
        409,
        410,
        411,
        412,
        413,
        414,
        415
      ],
      "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": 4,
    "instances": 12,
    "dataset_nodes": 6,
    "candidate_nodes": 84,
    "distinct_models": 17
  },
  "dataset_nodes": [
    {
      "node_id": "dataset_aed9252d40e94425",
      "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": 842901,
      "context_ref": "relations"
    },
    {
      "node_id": "dataset_536bf6fca69cc026",
      "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": 842902,
      "context_ref": "relations"
    },
    {
      "node_id": "dataset_cc2d85f228b64204",
      "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": 842903,
      "context_ref": "relations"
    },
    {
      "node_id": "dataset_57fed1289bc68187",
      "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": 842901,
      "context_ref": "shapes"
    },
    {
      "node_id": "dataset_16d796ce95ecf174",
      "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": 842902,
      "context_ref": "shapes"
    },
    {
      "node_id": "dataset_ca67f6ee4ab7c9fa",
      "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": 842903,
      "context_ref": "shapes"
    }
  ],
  "candidate_nodes": [
    {
      "node_id": "candidate_6dc2bfb1abd3acedb8ae",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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": 75,
        "total_samples_seen": 2400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          18,
          37,
          75
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 129700,
      "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010"
    },
    {
      "node_id": "candidate_1ff88bee20b5750886ae",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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": 75,
        "total_samples_seen": 2400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          18,
          37,
          75
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_984471b5ab216a29adf4",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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": 75,
        "total_samples_seen": 2400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          18,
          37,
          75
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_d370a44120e37327362b",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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_4dd55b3534f0e08d5749",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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_037ebf33a66bb98829ca",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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_8cbf0e4d791efb49a08a",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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": 75,
        "total_samples_seen": 2400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          18,
          37,
          75
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 129700,
      "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010"
    },
    {
      "node_id": "candidate_263a84dbee3415420f37",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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": 75,
        "total_samples_seen": 2400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          18,
          37,
          75
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_97cf98c643adaa13f038",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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": 75,
        "total_samples_seen": 2400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          18,
          37,
          75
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_b3a647c2f969e0104eca",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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_bf04bf97f220896826f4",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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_9111b4fd63f99a4619c6",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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_c64d9f879f32a436d0ef",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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": 75,
        "total_samples_seen": 2400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          18,
          37,
          75
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 129700,
      "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010"
    },
    {
      "node_id": "candidate_357dddb7394d60477312",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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": 75,
        "total_samples_seen": 2400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          18,
          37,
          75
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 16204,
      "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce"
    },
    {
      "node_id": "candidate_8fcda047626b8f22515f",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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": 75,
        "total_samples_seen": 2400,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_updates": 0,
        "batch_rng_device": "cpu",
        "paired_cpu_batch_stream": true,
        "persist_initial_state": true,
        "batch_seed_offset": 500000,
        "evaluation_split": "validation",
        "eval_steps": [
          18,
          37,
          75
        ],
        "record_training_diagnostics": true
      },
      "parameter_count": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_32181978996cabf895ed",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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_d3acbea5e15573e5b13c",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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_81e72f6daa6756ffcbfe",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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_dd21f46700a2ad1010f6",
      "dataset_node_id": "dataset_57fed1289bc68187",
      "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_c20af86a7e8d83f92bc7",
      "dataset_node_id": "dataset_57fed1289bc68187",
      "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_69b00e4721399251801e",
      "dataset_node_id": "dataset_57fed1289bc68187",
      "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_0193c02b28be3d71bc5c",
      "dataset_node_id": "dataset_57fed1289bc68187",
      "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": 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_93cfcaa885f9162ee227",
      "dataset_node_id": "dataset_57fed1289bc68187",
      "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_192902c5b48812986b23",
      "dataset_node_id": "dataset_57fed1289bc68187",
      "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_e3cfd913a3165d888dd4",
      "dataset_node_id": "dataset_57fed1289bc68187",
      "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": 0.003,
        "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": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_cf72091e51b277d32ab6",
      "dataset_node_id": "dataset_57fed1289bc68187",
      "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": 0.003,
        "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": 129700,
      "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010"
    },
    {
      "node_id": "candidate_5eba7433689d84096466",
      "dataset_node_id": "dataset_16d796ce95ecf174",
      "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_e73b093ae878fb579eb8",
      "dataset_node_id": "dataset_16d796ce95ecf174",
      "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_482310c10789d5287008",
      "dataset_node_id": "dataset_16d796ce95ecf174",
      "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_550b4e77af361f7a4464",
      "dataset_node_id": "dataset_16d796ce95ecf174",
      "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": 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_f6135d596acfca54cc75",
      "dataset_node_id": "dataset_16d796ce95ecf174",
      "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_826baae572159eab625c",
      "dataset_node_id": "dataset_16d796ce95ecf174",
      "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_54d5bf15f833498193fa",
      "dataset_node_id": "dataset_16d796ce95ecf174",
      "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": 0.003,
        "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": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_b5bc98e20dfafeefe261",
      "dataset_node_id": "dataset_16d796ce95ecf174",
      "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": 0.003,
        "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": 129700,
      "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010"
    },
    {
      "node_id": "candidate_bc967178393fba3243e2",
      "dataset_node_id": "dataset_ca67f6ee4ab7c9fa",
      "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_3c047d3664cd4b3c1c5e",
      "dataset_node_id": "dataset_ca67f6ee4ab7c9fa",
      "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_24443b0701b87898c365",
      "dataset_node_id": "dataset_ca67f6ee4ab7c9fa",
      "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_289d0a0358ab09a06844",
      "dataset_node_id": "dataset_ca67f6ee4ab7c9fa",
      "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": 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_f42e6dcd3bdb79dafb0e",
      "dataset_node_id": "dataset_ca67f6ee4ab7c9fa",
      "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_384affc445971835418c",
      "dataset_node_id": "dataset_ca67f6ee4ab7c9fa",
      "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_49335a90bc48e247ec66",
      "dataset_node_id": "dataset_ca67f6ee4ab7c9fa",
      "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": 0.003,
        "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": 19428,
      "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4"
    },
    {
      "node_id": "candidate_983fbd77986fead68406",
      "dataset_node_id": "dataset_ca67f6ee4ab7c9fa",
      "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": 0.003,
        "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": 129700,
      "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010"
    },
    {
      "node_id": "candidate_1c18fc797930667b5554",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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_cebf542b0c9dd216cf59",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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_8311a8ee675aef20c38e",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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": 12,
        "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": 26468,
      "model_code_sha256": "30bcd0cb7e95de286bc3f9a91c4aa654d2559794f6b4a451f4592b8630d0704e"
    },
    {
      "node_id": "candidate_6c37ff6172a94ef6e3af",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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_095f6edda493c046a96c",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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_269a8e9e175414391617",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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": 12,
        "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": 26596,
      "model_code_sha256": "b284df494e83dcd1a336fdb58dbb34eb5a06afb43c2f62169b659453848b54c4"
    },
    {
      "node_id": "candidate_21e3bc0b559aa842f774",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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_cffe94f10d775843b964",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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_966cbcd0109e11fe7624",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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": 12,
        "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": 26468,
      "model_code_sha256": "30bcd0cb7e95de286bc3f9a91c4aa654d2559794f6b4a451f4592b8630d0704e"
    },
    {
      "node_id": "candidate_6e80eb53d23db89710cf",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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_4dc751a92e6560f8ceca",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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_40656e7f962a4cdaebf1",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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": 12,
        "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": 26596,
      "model_code_sha256": "b284df494e83dcd1a336fdb58dbb34eb5a06afb43c2f62169b659453848b54c4"
    },
    {
      "node_id": "candidate_9413340d1808e35e38d2",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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_10057c486fc74e1cc8c3",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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_9b70caa50778b52c8b2b",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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": 12,
        "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": 26468,
      "model_code_sha256": "30bcd0cb7e95de286bc3f9a91c4aa654d2559794f6b4a451f4592b8630d0704e"
    },
    {
      "node_id": "candidate_8ec7408c41affba5ba69",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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_b0d33637a9a893bd3584",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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_50d950c767425f971cc0",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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": 12,
        "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": 26596,
      "model_code_sha256": "b284df494e83dcd1a336fdb58dbb34eb5a06afb43c2f62169b659453848b54c4"
    },
    {
      "node_id": "candidate_23e3443fc0d2d1dd9f3d",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": 46732,
      "model_code_sha256": "7183d61f2bf2188b283e14ce82318d20f5855fcd0a16bc4a57456c2081c89b67"
    },
    {
      "node_id": "candidate_51560fe5acf02a13457a",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 6,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": false,
        "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": 46732,
      "model_code_sha256": "6c1905d36843d749dd8b3ec752abc69e967eb1fe14a1ea0f4e6d5d0056f11ee0"
    },
    {
      "node_id": "candidate_cb27dff6653c4ab48187",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": 46732,
      "model_code_sha256": "7818c319584d6ddd071c6fa738ca50c22811f9f460d2e6e4469a54be52606813"
    },
    {
      "node_id": "candidate_0896cc8e3fabc2fe8c2d",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 6,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": false,
        "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": 46732,
      "model_code_sha256": "ebc1ab5fc558d25a5f9df7686b84ca5b3353795bfcdddd638b9b470c3e93b1c7"
    },
    {
      "node_id": "candidate_bba0f0d55d7892c14b30",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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_dad36112dfb7f7e146e1",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": false,
        "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": "fe876e8eaeb1b01ac4a21a337536fd164c10159bf1c5d049ea2715cca855718d"
    },
    {
      "node_id": "candidate_483f41cd6886f5e75f3b",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": "487f99a81970de56cd83e5ed249732dbfa7860e0a5389abfec6823e543153207"
    },
    {
      "node_id": "candidate_4927c746172e7414c383",
      "dataset_node_id": "dataset_aed9252d40e94425",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": false,
        "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": "1cc4fb9069b96900a53da98e555bc232d23fb2f3ce8a62596cb897b67dd82899"
    },
    {
      "node_id": "candidate_2376d6ec867863567e97",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": 46732,
      "model_code_sha256": "7183d61f2bf2188b283e14ce82318d20f5855fcd0a16bc4a57456c2081c89b67"
    },
    {
      "node_id": "candidate_ad3039dc2e5c6b8b2473",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 6,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": false,
        "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": 46732,
      "model_code_sha256": "6c1905d36843d749dd8b3ec752abc69e967eb1fe14a1ea0f4e6d5d0056f11ee0"
    },
    {
      "node_id": "candidate_e642aae4f07b02577827",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": 46732,
      "model_code_sha256": "7818c319584d6ddd071c6fa738ca50c22811f9f460d2e6e4469a54be52606813"
    },
    {
      "node_id": "candidate_8e07e57c4cfe33c7c341",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 6,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": false,
        "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": 46732,
      "model_code_sha256": "ebc1ab5fc558d25a5f9df7686b84ca5b3353795bfcdddd638b9b470c3e93b1c7"
    },
    {
      "node_id": "candidate_a213e5a7932138e3ea3b",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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_f40c302c8b5236763971",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": false,
        "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": "fe876e8eaeb1b01ac4a21a337536fd164c10159bf1c5d049ea2715cca855718d"
    },
    {
      "node_id": "candidate_82273fcd3d701b2bf0e0",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": "487f99a81970de56cd83e5ed249732dbfa7860e0a5389abfec6823e543153207"
    },
    {
      "node_id": "candidate_5bb6c4af2a5ff9b5ca27",
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": false,
        "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": "1cc4fb9069b96900a53da98e555bc232d23fb2f3ce8a62596cb897b67dd82899"
    },
    {
      "node_id": "candidate_9a90c18485205a946ae1",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": 46732,
      "model_code_sha256": "7183d61f2bf2188b283e14ce82318d20f5855fcd0a16bc4a57456c2081c89b67"
    },
    {
      "node_id": "candidate_dea57e6541351ba0d4e6",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 6,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": false,
        "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": 46732,
      "model_code_sha256": "6c1905d36843d749dd8b3ec752abc69e967eb1fe14a1ea0f4e6d5d0056f11ee0"
    },
    {
      "node_id": "candidate_38e568cc4cd7ed0123d8",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": 46732,
      "model_code_sha256": "7818c319584d6ddd071c6fa738ca50c22811f9f460d2e6e4469a54be52606813"
    },
    {
      "node_id": "candidate_ca6e32fc11a58f6d36e7",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 6,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": false,
        "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": 46732,
      "model_code_sha256": "ebc1ab5fc558d25a5f9df7686b84ca5b3353795bfcdddd638b9b470c3e93b1c7"
    },
    {
      "node_id": "candidate_fe7d3a88a6991af446d7",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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_0ec0b671b6d720c69814",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": true,
        "norm": "layer",
        "norm_first": false,
        "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": "fe876e8eaeb1b01ac4a21a337536fd164c10159bf1c5d049ea2715cca855718d"
    },
    {
      "node_id": "candidate_53b641693beab96ad7cd",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": "487f99a81970de56cd83e5ed249732dbfa7860e0a5389abfec6823e543153207"
    },
    {
      "node_id": "candidate_6cb23fbad1da49b99e05",
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "model": {
        "type": "image_classifier",
        "mode": "cnn",
        "width": 24,
        "depth": 2,
        "activation": "gelu",
        "residual": false,
        "norm": "layer",
        "norm_first": false,
        "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": 150,
        "total_samples_seen": 4800,
        "gradient_accumulation": 1,
        "clip_norm": 1.0,
        "schedule": "constant",
        "precision": "fp32",
        "tf32": false,
        "deterministic_algorithms": true,
        "warmup_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": "1cc4fb9069b96900a53da98e555bc232d23fb2f3ce8a62596cb897b67dd82899"
    }
  ],
  "model_validation_receipts": [
    {
      "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": "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": "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": "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": "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": "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": "ba29b76ad675b03faca5cfc19d952014e0fbc5a54c9a43dcb8de3537d629ca90",
      "parameters": 26468,
      "code_sha256": "30bcd0cb7e95de286bc3f9a91c4aa654d2559794f6b4a451f4592b8630d0704e",
      "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": "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": "9149b854524bd85c38092bfb42fc07fc5de168f86a386de212b88116f2bbe4d2",
      "parameters": 26596,
      "code_sha256": "b284df494e83dcd1a336fdb58dbb34eb5a06afb43c2f62169b659453848b54c4",
      "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": "96c059f918e64fff925bf16884bdcfc84d50bec925835a4751de85c3b99c90b8",
      "parameters": 46732,
      "code_sha256": "6c1905d36843d749dd8b3ec752abc69e967eb1fe14a1ea0f4e6d5d0056f11ee0",
      "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": "c976ac41bae7eba66c3d4fb83b784193e199ebfd41ab51dd7ea077fd83c8b4d4",
      "parameters": 46732,
      "code_sha256": "ebc1ab5fc558d25a5f9df7686b84ca5b3353795bfcdddd638b9b470c3e93b1c7",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    },
    {
      "model_sha256": "672404e722114bd793a31d04f40895970ba1d4b9a21dd20263542fc30520d23e",
      "parameters": 16204,
      "code_sha256": "fe876e8eaeb1b01ac4a21a337536fd164c10159bf1c5d049ea2715cca855718d",
      "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": "76fecb48314b66d03b5f0b2c227e22238096f0b7467fe69ccdae4949ee8ab8a0",
      "parameters": 16204,
      "code_sha256": "1cc4fb9069b96900a53da98e555bc232d23fb2f3ce8a62596cb897b67dd82899",
      "smoke": "standalone_render_forward_backward_and_canonical_parity",
      "output_shape": [
        2,
        4
      ],
      "canonical_max_abs_error": 0.0
    }
  ],
  "templates": [
    {
      "id": "operator_learning_speed_s842901",
      "category": "operator_budget",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.mode"
      ],
      "hypothesis": "Architecture recipes can learn at different speeds; rankings may depend on the finite optimization horizon.",
      "explanation_target": "Compare all three recipes at both 75 and 300 updates. A 75-step ranking does not establish asymptotic architecture quality. Width, parameterization, pooling, patching and positional representation all differ in the operator recipe; recorded actual fields make the bundle explicit.",
      "controls": [
        "Same fixed materialized train/validation/test/OOD splits within each context and data seed.",
        "Each primary triple has equal updates and batch size; diagnostic horizons are explicitly separate.",
        "All unlisted fields inherit explicit defaults; measured parameter counts and all actual changed fields are exposed.",
        "No matched-parameter or matched-FLOP claim; mode comparisons are bundled recipes.",
        "Final checkpoints only; fixed horizon, no early stopping or validation-selected checkpoint."
      ],
      "shortcut_checks": [
        "globally stronger operator always wins",
        "larger parameter count always wins",
        "short-horizon order persists at 300 updates"
      ],
      "admission": "Freeze ascending-mean validation directions after four discovery seeds; admit only when both adjacent paired sign tests pass the independent 24-hypothesis Holm family and each observed mean test-CE gap exceeds 0.005. No replacement after failure.",
      "template_id": "operator_learning_speed",
      "instance_seed": 842901,
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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": {
          "mlp_75": {
            "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
          },
          "cnn_75": {
            "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
          },
          "vit_75": {
            "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": "mlp_75",
          "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": 75,
            "total_samples_seen": 2400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              18,
              37,
              75
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 129700,
          "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010",
          "node_id": "candidate_6dc2bfb1abd3acedb8ae"
        },
        {
          "id": "cnn_75",
          "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": 75,
            "total_samples_seen": 2400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              18,
              37,
              75
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_1ff88bee20b5750886ae"
        },
        {
          "id": "vit_75",
          "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": 75,
            "total_samples_seen": 2400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              18,
              37,
              75
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_984471b5ab216a29adf4"
        }
      ],
      "diagnostics": [
        {
          "id": "mlp_300",
          "role": "diagnostic",
          "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_d370a44120e37327362b"
        },
        {
          "id": "cnn_300",
          "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_4dd55b3534f0e08d5749"
        },
        {
          "id": "vit_300",
          "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": 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_037ebf33a66bb98829ca"
        }
      ]
    },
    {
      "id": "operator_learning_speed_s842902",
      "category": "operator_budget",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.mode"
      ],
      "hypothesis": "Architecture recipes can learn at different speeds; rankings may depend on the finite optimization horizon.",
      "explanation_target": "Compare all three recipes at both 75 and 300 updates. A 75-step ranking does not establish asymptotic architecture quality. Width, parameterization, pooling, patching and positional representation all differ in the operator recipe; recorded actual fields make the bundle explicit.",
      "controls": [
        "Same fixed materialized train/validation/test/OOD splits within each context and data seed.",
        "Each primary triple has equal updates and batch size; diagnostic horizons are explicitly separate.",
        "All unlisted fields inherit explicit defaults; measured parameter counts and all actual changed fields are exposed.",
        "No matched-parameter or matched-FLOP claim; mode comparisons are bundled recipes.",
        "Final checkpoints only; fixed horizon, no early stopping or validation-selected checkpoint."
      ],
      "shortcut_checks": [
        "globally stronger operator always wins",
        "larger parameter count always wins",
        "short-horizon order persists at 300 updates"
      ],
      "admission": "Freeze ascending-mean validation directions after four discovery seeds; admit only when both adjacent paired sign tests pass the independent 24-hypothesis Holm family and each observed mean test-CE gap exceeds 0.005. No replacement after failure.",
      "template_id": "operator_learning_speed",
      "instance_seed": 842902,
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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": {
          "mlp_75": {
            "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
          },
          "cnn_75": {
            "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
          },
          "vit_75": {
            "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": "mlp_75",
          "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": 75,
            "total_samples_seen": 2400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              18,
              37,
              75
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 129700,
          "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010",
          "node_id": "candidate_8cbf0e4d791efb49a08a"
        },
        {
          "id": "cnn_75",
          "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": 75,
            "total_samples_seen": 2400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              18,
              37,
              75
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_263a84dbee3415420f37"
        },
        {
          "id": "vit_75",
          "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": 75,
            "total_samples_seen": 2400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              18,
              37,
              75
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_97cf98c643adaa13f038"
        }
      ],
      "diagnostics": [
        {
          "id": "mlp_300",
          "role": "diagnostic",
          "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_b3a647c2f969e0104eca"
        },
        {
          "id": "cnn_300",
          "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_bf04bf97f220896826f4"
        },
        {
          "id": "vit_300",
          "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": 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_9111b4fd63f99a4619c6"
        }
      ]
    },
    {
      "id": "operator_learning_speed_s842903",
      "category": "operator_budget",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.mode"
      ],
      "hypothesis": "Architecture recipes can learn at different speeds; rankings may depend on the finite optimization horizon.",
      "explanation_target": "Compare all three recipes at both 75 and 300 updates. A 75-step ranking does not establish asymptotic architecture quality. Width, parameterization, pooling, patching and positional representation all differ in the operator recipe; recorded actual fields make the bundle explicit.",
      "controls": [
        "Same fixed materialized train/validation/test/OOD splits within each context and data seed.",
        "Each primary triple has equal updates and batch size; diagnostic horizons are explicitly separate.",
        "All unlisted fields inherit explicit defaults; measured parameter counts and all actual changed fields are exposed.",
        "No matched-parameter or matched-FLOP claim; mode comparisons are bundled recipes.",
        "Final checkpoints only; fixed horizon, no early stopping or validation-selected checkpoint."
      ],
      "shortcut_checks": [
        "globally stronger operator always wins",
        "larger parameter count always wins",
        "short-horizon order persists at 300 updates"
      ],
      "admission": "Freeze ascending-mean validation directions after four discovery seeds; admit only when both adjacent paired sign tests pass the independent 24-hypothesis Holm family and each observed mean test-CE gap exceeds 0.005. No replacement after failure.",
      "template_id": "operator_learning_speed",
      "instance_seed": 842903,
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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": {
          "mlp_75": {
            "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
          },
          "cnn_75": {
            "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
          },
          "vit_75": {
            "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": "mlp_75",
          "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": 75,
            "total_samples_seen": 2400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              18,
              37,
              75
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 129700,
          "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010",
          "node_id": "candidate_c64d9f879f32a436d0ef"
        },
        {
          "id": "cnn_75",
          "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": 75,
            "total_samples_seen": 2400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              18,
              37,
              75
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 16204,
          "model_code_sha256": "4ef66495d59b36a884f7ef41825e8e200faeeebeec160a8962a4541f1c6d6bce",
          "node_id": "candidate_357dddb7394d60477312"
        },
        {
          "id": "vit_75",
          "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": 75,
            "total_samples_seen": 2400,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_updates": 0,
            "batch_rng_device": "cpu",
            "paired_cpu_batch_stream": true,
            "persist_initial_state": true,
            "batch_seed_offset": 500000,
            "evaluation_split": "validation",
            "eval_steps": [
              18,
              37,
              75
            ],
            "record_training_diagnostics": true
          },
          "parameter_count": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_8fcda047626b8f22515f"
        }
      ],
      "diagnostics": [
        {
          "id": "mlp_300",
          "role": "diagnostic",
          "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_32181978996cabf895ed"
        },
        {
          "id": "cnn_300",
          "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_d3acbea5e15573e5b13c"
        },
        {
          "id": "vit_300",
          "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": 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_81e72f6daa6756ffcbfe"
        }
      ]
    },
    {
      "id": "operator_shape_generalization_s842901",
      "category": "operator_generalization",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.mode",
        "optimizer.lr"
      ],
      "hypothesis": "Operator-specific finite-data learning and generalization may differ, while learning rate and CNN readout can alter the observed recipe gap.",
      "explanation_target": "Use the complete CNN pooling x learning-rate control and both learning rates for ViT and MLP. Compare final complete training CE, validation trajectories and held-out CE before distinguishing optimization failure from overfitting. Do not attribute the bundled primary comparison solely to architecture or extrapolate to natural images.",
      "controls": [
        "Same fixed materialized train/validation/test/OOD splits within each context and data seed.",
        "Each primary triple has equal updates and batch size; diagnostic horizons are explicitly separate.",
        "All unlisted fields inherit explicit defaults; measured parameter counts and all actual changed fields are exposed.",
        "No matched-parameter or matched-FLOP claim; mode comparisons are bundled recipes.",
        "Final checkpoints only; fixed horizon, no early stopping or validation-selected checkpoint."
      ],
      "shortcut_checks": [
        "Transformer always wins image tasks",
        "lowest learning rate always generalizes best",
        "training fit alone predicts held-out cross-entropy"
      ],
      "admission": "Freeze ascending-mean validation directions after four discovery seeds; admit only when both adjacent paired sign tests pass the independent 24-hypothesis Holm family and each observed mean test-CE gap exceeds 0.005. No replacement after failure.",
      "template_id": "operator_shape_generalization",
      "instance_seed": 842901,
      "dataset_node_id": "dataset_57fed1289bc68187",
      "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"
        ],
        "comparison_class": "bundled_recipe",
        "field_values": {
          "cnn_spatial_lr003": {
            "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,
            "optimizer.lr": 0.003
          },
          "vit_lr001": {
            "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
          },
          "mlp_lr001": {
            "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
          }
        },
        "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_spatial_lr003",
          "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": 0.003,
            "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_dd21f46700a2ad1010f6"
        },
        {
          "id": "vit_lr001",
          "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_c20af86a7e8d83f92bc7"
        },
        {
          "id": "mlp_lr001",
          "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_69b00e4721399251801e"
        }
      ],
      "diagnostics": [
        {
          "id": "cnn_spatial_lr001",
          "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": 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_0193c02b28be3d71bc5c"
        },
        {
          "id": "cnn_mean_lr001",
          "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": 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_93cfcaa885f9162ee227"
        },
        {
          "id": "cnn_mean_lr003",
          "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": 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_192902c5b48812986b23"
        },
        {
          "id": "vit_lr003",
          "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": 0.003,
            "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": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_e3cfd913a3165d888dd4"
        },
        {
          "id": "mlp_lr003",
          "role": "diagnostic",
          "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": 0.003,
            "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": 129700,
          "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010",
          "node_id": "candidate_cf72091e51b277d32ab6"
        }
      ]
    },
    {
      "id": "operator_shape_generalization_s842902",
      "category": "operator_generalization",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.mode",
        "optimizer.lr"
      ],
      "hypothesis": "Operator-specific finite-data learning and generalization may differ, while learning rate and CNN readout can alter the observed recipe gap.",
      "explanation_target": "Use the complete CNN pooling x learning-rate control and both learning rates for ViT and MLP. Compare final complete training CE, validation trajectories and held-out CE before distinguishing optimization failure from overfitting. Do not attribute the bundled primary comparison solely to architecture or extrapolate to natural images.",
      "controls": [
        "Same fixed materialized train/validation/test/OOD splits within each context and data seed.",
        "Each primary triple has equal updates and batch size; diagnostic horizons are explicitly separate.",
        "All unlisted fields inherit explicit defaults; measured parameter counts and all actual changed fields are exposed.",
        "No matched-parameter or matched-FLOP claim; mode comparisons are bundled recipes.",
        "Final checkpoints only; fixed horizon, no early stopping or validation-selected checkpoint."
      ],
      "shortcut_checks": [
        "Transformer always wins image tasks",
        "lowest learning rate always generalizes best",
        "training fit alone predicts held-out cross-entropy"
      ],
      "admission": "Freeze ascending-mean validation directions after four discovery seeds; admit only when both adjacent paired sign tests pass the independent 24-hypothesis Holm family and each observed mean test-CE gap exceeds 0.005. No replacement after failure.",
      "template_id": "operator_shape_generalization",
      "instance_seed": 842902,
      "dataset_node_id": "dataset_16d796ce95ecf174",
      "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"
        ],
        "comparison_class": "bundled_recipe",
        "field_values": {
          "cnn_spatial_lr003": {
            "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,
            "optimizer.lr": 0.003
          },
          "vit_lr001": {
            "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
          },
          "mlp_lr001": {
            "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
          }
        },
        "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_spatial_lr003",
          "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": 0.003,
            "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_5eba7433689d84096466"
        },
        {
          "id": "vit_lr001",
          "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_e73b093ae878fb579eb8"
        },
        {
          "id": "mlp_lr001",
          "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_482310c10789d5287008"
        }
      ],
      "diagnostics": [
        {
          "id": "cnn_spatial_lr001",
          "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": 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_550b4e77af361f7a4464"
        },
        {
          "id": "cnn_mean_lr001",
          "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": 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_f6135d596acfca54cc75"
        },
        {
          "id": "cnn_mean_lr003",
          "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": 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_826baae572159eab625c"
        },
        {
          "id": "vit_lr003",
          "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": 0.003,
            "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": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_54d5bf15f833498193fa"
        },
        {
          "id": "mlp_lr003",
          "role": "diagnostic",
          "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": 0.003,
            "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": 129700,
          "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010",
          "node_id": "candidate_b5bc98e20dfafeefe261"
        }
      ]
    },
    {
      "id": "operator_shape_generalization_s842903",
      "category": "operator_generalization",
      "context_ref": "shapes",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.mode",
        "optimizer.lr"
      ],
      "hypothesis": "Operator-specific finite-data learning and generalization may differ, while learning rate and CNN readout can alter the observed recipe gap.",
      "explanation_target": "Use the complete CNN pooling x learning-rate control and both learning rates for ViT and MLP. Compare final complete training CE, validation trajectories and held-out CE before distinguishing optimization failure from overfitting. Do not attribute the bundled primary comparison solely to architecture or extrapolate to natural images.",
      "controls": [
        "Same fixed materialized train/validation/test/OOD splits within each context and data seed.",
        "Each primary triple has equal updates and batch size; diagnostic horizons are explicitly separate.",
        "All unlisted fields inherit explicit defaults; measured parameter counts and all actual changed fields are exposed.",
        "No matched-parameter or matched-FLOP claim; mode comparisons are bundled recipes.",
        "Final checkpoints only; fixed horizon, no early stopping or validation-selected checkpoint."
      ],
      "shortcut_checks": [
        "Transformer always wins image tasks",
        "lowest learning rate always generalizes best",
        "training fit alone predicts held-out cross-entropy"
      ],
      "admission": "Freeze ascending-mean validation directions after four discovery seeds; admit only when both adjacent paired sign tests pass the independent 24-hypothesis Holm family and each observed mean test-CE gap exceeds 0.005. No replacement after failure.",
      "template_id": "operator_shape_generalization",
      "instance_seed": 842903,
      "dataset_node_id": "dataset_ca67f6ee4ab7c9fa",
      "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"
        ],
        "comparison_class": "bundled_recipe",
        "field_values": {
          "cnn_spatial_lr003": {
            "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,
            "optimizer.lr": 0.003
          },
          "vit_lr001": {
            "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
          },
          "mlp_lr001": {
            "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
          }
        },
        "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_spatial_lr003",
          "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": 0.003,
            "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_bc967178393fba3243e2"
        },
        {
          "id": "vit_lr001",
          "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_3c047d3664cd4b3c1c5e"
        },
        {
          "id": "mlp_lr001",
          "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_24443b0701b87898c365"
        }
      ],
      "diagnostics": [
        {
          "id": "cnn_spatial_lr001",
          "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": 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_289d0a0358ab09a06844"
        },
        {
          "id": "cnn_mean_lr001",
          "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": 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_f42e6dcd3bdb79dafb0e"
        },
        {
          "id": "cnn_mean_lr003",
          "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": 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_384affc445971835418c"
        },
        {
          "id": "vit_lr003",
          "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": 0.003,
            "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": 19428,
          "model_code_sha256": "430cca84c030f92c02ccd4e9590c8d84e9e15cf269399565255af3c58c15afb4",
          "node_id": "candidate_49335a90bc48e247ec66"
        },
        {
          "id": "mlp_lr003",
          "role": "diagnostic",
          "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": 0.003,
            "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": 129700,
          "model_code_sha256": "adfddaed3cfdb887611a2e75ef156781e96c09289c5eff9b0f79b5847f0c0010",
          "node_id": "candidate_983fbd77986fead68406"
        }
      ]
    },
    {
      "id": "patch_coordinate_retention_s842901",
      "category": "patch_position",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.patch_size"
      ],
      "hypothesis": "Patch size changes both tokenization and retained within-patch positional structure; explicit positional embeddings may interact with this trade-off.",
      "explanation_target": "Use the complete patch-size x position-embedding control. Removing learned position embeddings leaves ordered within-patch pixels inside the linear projection. Larger patches change token count, projection parameters and optimization as well as spatial aggregation; patch projection is not a guaranteed antialiasing filter. New p12 cells are untested design hypotheses and may fail admission.",
      "controls": [
        "Same fixed materialized train/validation/test/OOD splits within each context and data seed.",
        "Each primary triple has equal updates and batch size; diagnostic horizons are explicitly separate.",
        "All unlisted fields inherit explicit defaults; measured parameter counts and all actual changed fields are exposed.",
        "No matched-parameter or matched-FLOP claim; mode comparisons are bundled recipes.",
        "Final checkpoints only; fixed horizon, no early stopping or validation-selected checkpoint."
      ],
      "shortcut_checks": [
        "no positional embeddings makes all position learning impossible",
        "smaller patches always preserve more useful information",
        "patch projection guarantees antialiasing"
      ],
      "admission": "Freeze ascending-mean validation directions after four discovery seeds; admit only when both adjacent paired sign tests pass the independent 24-hypothesis Holm family and each observed mean test-CE gap exceeds 0.005. No replacement after failure.",
      "template_id": "patch_coordinate_retention",
      "instance_seed": 842901,
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "p3_none": {
            "model.patch_size": 3
          },
          "p6_none": {
            "model.patch_size": 6
          },
          "p12_none": {
            "model.patch_size": 12
          }
        },
        "actual_changed_fields": [
          "model.patch_size"
        ]
      },
      "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_1c18fc797930667b5554"
        },
        {
          "id": "p6_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": 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_cebf542b0c9dd216cf59"
        },
        {
          "id": "p12_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": 12,
            "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": 26468,
          "model_code_sha256": "30bcd0cb7e95de286bc3f9a91c4aa654d2559794f6b4a451f4592b8630d0704e",
          "node_id": "candidate_8311a8ee675aef20c38e"
        }
      ],
      "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_6c37ff6172a94ef6e3af"
        },
        {
          "id": "p6_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": 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_095f6edda493c046a96c"
        },
        {
          "id": "p12_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": 12,
            "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": 26596,
          "model_code_sha256": "b284df494e83dcd1a336fdb58dbb34eb5a06afb43c2f62169b659453848b54c4",
          "node_id": "candidate_269a8e9e175414391617"
        }
      ]
    },
    {
      "id": "patch_coordinate_retention_s842902",
      "category": "patch_position",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.patch_size"
      ],
      "hypothesis": "Patch size changes both tokenization and retained within-patch positional structure; explicit positional embeddings may interact with this trade-off.",
      "explanation_target": "Use the complete patch-size x position-embedding control. Removing learned position embeddings leaves ordered within-patch pixels inside the linear projection. Larger patches change token count, projection parameters and optimization as well as spatial aggregation; patch projection is not a guaranteed antialiasing filter. New p12 cells are untested design hypotheses and may fail admission.",
      "controls": [
        "Same fixed materialized train/validation/test/OOD splits within each context and data seed.",
        "Each primary triple has equal updates and batch size; diagnostic horizons are explicitly separate.",
        "All unlisted fields inherit explicit defaults; measured parameter counts and all actual changed fields are exposed.",
        "No matched-parameter or matched-FLOP claim; mode comparisons are bundled recipes.",
        "Final checkpoints only; fixed horizon, no early stopping or validation-selected checkpoint."
      ],
      "shortcut_checks": [
        "no positional embeddings makes all position learning impossible",
        "smaller patches always preserve more useful information",
        "patch projection guarantees antialiasing"
      ],
      "admission": "Freeze ascending-mean validation directions after four discovery seeds; admit only when both adjacent paired sign tests pass the independent 24-hypothesis Holm family and each observed mean test-CE gap exceeds 0.005. No replacement after failure.",
      "template_id": "patch_coordinate_retention",
      "instance_seed": 842902,
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "p3_none": {
            "model.patch_size": 3
          },
          "p6_none": {
            "model.patch_size": 6
          },
          "p12_none": {
            "model.patch_size": 12
          }
        },
        "actual_changed_fields": [
          "model.patch_size"
        ]
      },
      "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_21e3bc0b559aa842f774"
        },
        {
          "id": "p6_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": 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_cffe94f10d775843b964"
        },
        {
          "id": "p12_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": 12,
            "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": 26468,
          "model_code_sha256": "30bcd0cb7e95de286bc3f9a91c4aa654d2559794f6b4a451f4592b8630d0704e",
          "node_id": "candidate_966cbcd0109e11fe7624"
        }
      ],
      "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_6e80eb53d23db89710cf"
        },
        {
          "id": "p6_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": 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_4dc751a92e6560f8ceca"
        },
        {
          "id": "p12_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": 12,
            "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": 26596,
          "model_code_sha256": "b284df494e83dcd1a336fdb58dbb34eb5a06afb43c2f62169b659453848b54c4",
          "node_id": "candidate_40656e7f962a4cdaebf1"
        }
      ]
    },
    {
      "id": "patch_coordinate_retention_s842903",
      "category": "patch_position",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.patch_size"
      ],
      "hypothesis": "Patch size changes both tokenization and retained within-patch positional structure; explicit positional embeddings may interact with this trade-off.",
      "explanation_target": "Use the complete patch-size x position-embedding control. Removing learned position embeddings leaves ordered within-patch pixels inside the linear projection. Larger patches change token count, projection parameters and optimization as well as spatial aggregation; patch projection is not a guaranteed antialiasing filter. New p12 cells are untested design hypotheses and may fail admission.",
      "controls": [
        "Same fixed materialized train/validation/test/OOD splits within each context and data seed.",
        "Each primary triple has equal updates and batch size; diagnostic horizons are explicitly separate.",
        "All unlisted fields inherit explicit defaults; measured parameter counts and all actual changed fields are exposed.",
        "No matched-parameter or matched-FLOP claim; mode comparisons are bundled recipes.",
        "Final checkpoints only; fixed horizon, no early stopping or validation-selected checkpoint."
      ],
      "shortcut_checks": [
        "no positional embeddings makes all position learning impossible",
        "smaller patches always preserve more useful information",
        "patch projection guarantees antialiasing"
      ],
      "admission": "Freeze ascending-mean validation directions after four discovery seeds; admit only when both adjacent paired sign tests pass the independent 24-hypothesis Holm family and each observed mean test-CE gap exceeds 0.005. No replacement after failure.",
      "template_id": "patch_coordinate_retention",
      "instance_seed": 842903,
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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"
        ],
        "undeclared_fields": [],
        "comparison_class": "single_factor",
        "field_values": {
          "p3_none": {
            "model.patch_size": 3
          },
          "p6_none": {
            "model.patch_size": 6
          },
          "p12_none": {
            "model.patch_size": 12
          }
        },
        "actual_changed_fields": [
          "model.patch_size"
        ]
      },
      "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_9413340d1808e35e38d2"
        },
        {
          "id": "p6_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": 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_10057c486fc74e1cc8c3"
        },
        {
          "id": "p12_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": 12,
            "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": 26468,
          "model_code_sha256": "30bcd0cb7e95de286bc3f9a91c4aa654d2559794f6b4a451f4592b8630d0704e",
          "node_id": "candidate_9b70caa50778b52c8b2b"
        }
      ],
      "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_8ec7408c41affba5ba69"
        },
        {
          "id": "p6_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": 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_b0d33637a9a893bd3584"
        },
        {
          "id": "p12_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": 12,
            "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": 26596,
          "model_code_sha256": "b284df494e83dcd1a336fdb58dbb34eb5a06afb43c2f62169b659453848b54c4",
          "node_id": "candidate_50d950c767425f971cc0"
        }
      ]
    },
    {
      "id": "residual_norm_placement_s842901",
      "category": "residual_normalization",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.residual",
        "model.norm_first"
      ],
      "hypothesis": "Residual paths and normalization placement can interact with depth under the same finite update budget.",
      "explanation_target": "Use all eight depth x residual x normalization-placement cells. CNN LayerNorm is per spatial location across channels. Pre/post placement changes the forward function as implemented, not just a label. Global gradient norms and clipping are nonexclusive diagnostics and do not establish layerwise vanishing or exploding gradients. Post-norm cells are untested and may fail admission.",
      "controls": [
        "Same fixed materialized train/validation/test/OOD splits within each context and data seed.",
        "Each primary triple has equal updates and batch size; diagnostic horizons are explicitly separate.",
        "All unlisted fields inherit explicit defaults; measured parameter counts and all actual changed fields are exposed.",
        "No matched-parameter or matched-FLOP claim; mode comparisons are bundled recipes.",
        "Final checkpoints only; fixed horizon, no early stopping or validation-selected checkpoint."
      ],
      "shortcut_checks": [
        "residual paths guarantee success",
        "pre-norm always wins",
        "nonzero global gradient rules out layerwise vanishing"
      ],
      "admission": "Freeze ascending-mean validation directions after four discovery seeds; admit only when both adjacent paired sign tests pass the independent 24-hypothesis Holm family and each observed mean test-CE gap exceeds 0.005. No replacement after failure.",
      "template_id": "residual_norm_placement",
      "instance_seed": 842901,
      "dataset_node_id": "dataset_aed9252d40e94425",
      "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.norm_first",
          "model.residual"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "d6_residual_pre": {
            "model.norm_first": true,
            "model.residual": true
          },
          "d6_residual_post": {
            "model.norm_first": false,
            "model.residual": true
          },
          "d6_plain_pre": {
            "model.norm_first": true,
            "model.residual": false
          }
        },
        "actual_changed_fields": [
          "model.norm_first",
          "model.residual"
        ]
      },
      "choices": [
        {
          "id": "d6_residual_pre",
          "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": 46732,
          "model_code_sha256": "7183d61f2bf2188b283e14ce82318d20f5855fcd0a16bc4a57456c2081c89b67",
          "node_id": "candidate_23e3443fc0d2d1dd9f3d"
        },
        {
          "id": "d6_residual_post",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 6,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": false,
            "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": 46732,
          "model_code_sha256": "6c1905d36843d749dd8b3ec752abc69e967eb1fe14a1ea0f4e6d5d0056f11ee0",
          "node_id": "candidate_51560fe5acf02a13457a"
        },
        {
          "id": "d6_plain_pre",
          "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": 46732,
          "model_code_sha256": "7818c319584d6ddd071c6fa738ca50c22811f9f460d2e6e4469a54be52606813",
          "node_id": "candidate_cb27dff6653c4ab48187"
        }
      ],
      "diagnostics": [
        {
          "id": "d6_plain_post",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 6,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": false,
            "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": 46732,
          "model_code_sha256": "ebc1ab5fc558d25a5f9df7686b84ca5b3353795bfcdddd638b9b470c3e93b1c7",
          "node_id": "candidate_0896cc8e3fabc2fe8c2d"
        },
        {
          "id": "d2_residual_pre",
          "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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_bba0f0d55d7892c14b30"
        },
        {
          "id": "d2_residual_post",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": false,
            "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": "fe876e8eaeb1b01ac4a21a337536fd164c10159bf1c5d049ea2715cca855718d",
          "node_id": "candidate_dad36112dfb7f7e146e1"
        },
        {
          "id": "d2_plain_pre",
          "role": "diagnostic",
          "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": "487f99a81970de56cd83e5ed249732dbfa7860e0a5389abfec6823e543153207",
          "node_id": "candidate_483f41cd6886f5e75f3b"
        },
        {
          "id": "d2_plain_post",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": false,
            "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": "1cc4fb9069b96900a53da98e555bc232d23fb2f3ce8a62596cb897b67dd82899",
          "node_id": "candidate_4927c746172e7414c383"
        }
      ]
    },
    {
      "id": "residual_norm_placement_s842902",
      "category": "residual_normalization",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.residual",
        "model.norm_first"
      ],
      "hypothesis": "Residual paths and normalization placement can interact with depth under the same finite update budget.",
      "explanation_target": "Use all eight depth x residual x normalization-placement cells. CNN LayerNorm is per spatial location across channels. Pre/post placement changes the forward function as implemented, not just a label. Global gradient norms and clipping are nonexclusive diagnostics and do not establish layerwise vanishing or exploding gradients. Post-norm cells are untested and may fail admission.",
      "controls": [
        "Same fixed materialized train/validation/test/OOD splits within each context and data seed.",
        "Each primary triple has equal updates and batch size; diagnostic horizons are explicitly separate.",
        "All unlisted fields inherit explicit defaults; measured parameter counts and all actual changed fields are exposed.",
        "No matched-parameter or matched-FLOP claim; mode comparisons are bundled recipes.",
        "Final checkpoints only; fixed horizon, no early stopping or validation-selected checkpoint."
      ],
      "shortcut_checks": [
        "residual paths guarantee success",
        "pre-norm always wins",
        "nonzero global gradient rules out layerwise vanishing"
      ],
      "admission": "Freeze ascending-mean validation directions after four discovery seeds; admit only when both adjacent paired sign tests pass the independent 24-hypothesis Holm family and each observed mean test-CE gap exceeds 0.005. No replacement after failure.",
      "template_id": "residual_norm_placement",
      "instance_seed": 842902,
      "dataset_node_id": "dataset_536bf6fca69cc026",
      "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.norm_first",
          "model.residual"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "d6_residual_pre": {
            "model.norm_first": true,
            "model.residual": true
          },
          "d6_residual_post": {
            "model.norm_first": false,
            "model.residual": true
          },
          "d6_plain_pre": {
            "model.norm_first": true,
            "model.residual": false
          }
        },
        "actual_changed_fields": [
          "model.norm_first",
          "model.residual"
        ]
      },
      "choices": [
        {
          "id": "d6_residual_pre",
          "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": 46732,
          "model_code_sha256": "7183d61f2bf2188b283e14ce82318d20f5855fcd0a16bc4a57456c2081c89b67",
          "node_id": "candidate_2376d6ec867863567e97"
        },
        {
          "id": "d6_residual_post",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 6,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": false,
            "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": 46732,
          "model_code_sha256": "6c1905d36843d749dd8b3ec752abc69e967eb1fe14a1ea0f4e6d5d0056f11ee0",
          "node_id": "candidate_ad3039dc2e5c6b8b2473"
        },
        {
          "id": "d6_plain_pre",
          "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": 46732,
          "model_code_sha256": "7818c319584d6ddd071c6fa738ca50c22811f9f460d2e6e4469a54be52606813",
          "node_id": "candidate_e642aae4f07b02577827"
        }
      ],
      "diagnostics": [
        {
          "id": "d6_plain_post",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 6,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": false,
            "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": 46732,
          "model_code_sha256": "ebc1ab5fc558d25a5f9df7686b84ca5b3353795bfcdddd638b9b470c3e93b1c7",
          "node_id": "candidate_8e07e57c4cfe33c7c341"
        },
        {
          "id": "d2_residual_pre",
          "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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_a213e5a7932138e3ea3b"
        },
        {
          "id": "d2_residual_post",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": false,
            "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": "fe876e8eaeb1b01ac4a21a337536fd164c10159bf1c5d049ea2715cca855718d",
          "node_id": "candidate_f40c302c8b5236763971"
        },
        {
          "id": "d2_plain_pre",
          "role": "diagnostic",
          "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": "487f99a81970de56cd83e5ed249732dbfa7860e0a5389abfec6823e543153207",
          "node_id": "candidate_82273fcd3d701b2bf0e0"
        },
        {
          "id": "d2_plain_post",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": false,
            "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": "1cc4fb9069b96900a53da98e555bc232d23fb2f3ce8a62596cb897b67dd82899",
          "node_id": "candidate_5bb6c4af2a5ff9b5ca27"
        }
      ]
    },
    {
      "id": "residual_norm_placement_s842903",
      "category": "residual_normalization",
      "context_ref": "relations",
      "budget_semantics": "matched_steps",
      "varying_axes": [
        "model.residual",
        "model.norm_first"
      ],
      "hypothesis": "Residual paths and normalization placement can interact with depth under the same finite update budget.",
      "explanation_target": "Use all eight depth x residual x normalization-placement cells. CNN LayerNorm is per spatial location across channels. Pre/post placement changes the forward function as implemented, not just a label. Global gradient norms and clipping are nonexclusive diagnostics and do not establish layerwise vanishing or exploding gradients. Post-norm cells are untested and may fail admission.",
      "controls": [
        "Same fixed materialized train/validation/test/OOD splits within each context and data seed.",
        "Each primary triple has equal updates and batch size; diagnostic horizons are explicitly separate.",
        "All unlisted fields inherit explicit defaults; measured parameter counts and all actual changed fields are exposed.",
        "No matched-parameter or matched-FLOP claim; mode comparisons are bundled recipes.",
        "Final checkpoints only; fixed horizon, no early stopping or validation-selected checkpoint."
      ],
      "shortcut_checks": [
        "residual paths guarantee success",
        "pre-norm always wins",
        "nonzero global gradient rules out layerwise vanishing"
      ],
      "admission": "Freeze ascending-mean validation directions after four discovery seeds; admit only when both adjacent paired sign tests pass the independent 24-hypothesis Holm family and each observed mean test-CE gap exceeds 0.005. No replacement after failure.",
      "template_id": "residual_norm_placement",
      "instance_seed": 842903,
      "dataset_node_id": "dataset_cc2d85f228b64204",
      "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.norm_first",
          "model.residual"
        ],
        "undeclared_fields": [],
        "comparison_class": "factorial",
        "field_values": {
          "d6_residual_pre": {
            "model.norm_first": true,
            "model.residual": true
          },
          "d6_residual_post": {
            "model.norm_first": false,
            "model.residual": true
          },
          "d6_plain_pre": {
            "model.norm_first": true,
            "model.residual": false
          }
        },
        "actual_changed_fields": [
          "model.norm_first",
          "model.residual"
        ]
      },
      "choices": [
        {
          "id": "d6_residual_pre",
          "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": 46732,
          "model_code_sha256": "7183d61f2bf2188b283e14ce82318d20f5855fcd0a16bc4a57456c2081c89b67",
          "node_id": "candidate_9a90c18485205a946ae1"
        },
        {
          "id": "d6_residual_post",
          "role": "primary",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 6,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": false,
            "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": 46732,
          "model_code_sha256": "6c1905d36843d749dd8b3ec752abc69e967eb1fe14a1ea0f4e6d5d0056f11ee0",
          "node_id": "candidate_dea57e6541351ba0d4e6"
        },
        {
          "id": "d6_plain_pre",
          "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": 46732,
          "model_code_sha256": "7818c319584d6ddd071c6fa738ca50c22811f9f460d2e6e4469a54be52606813",
          "node_id": "candidate_38e568cc4cd7ed0123d8"
        }
      ],
      "diagnostics": [
        {
          "id": "d6_plain_post",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 6,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": false,
            "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": 46732,
          "model_code_sha256": "ebc1ab5fc558d25a5f9df7686b84ca5b3353795bfcdddd638b9b470c3e93b1c7",
          "node_id": "candidate_ca6e32fc11a58f6d36e7"
        },
        {
          "id": "d2_residual_pre",
          "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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_fe7d3a88a6991af446d7"
        },
        {
          "id": "d2_residual_post",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": true,
            "norm": "layer",
            "norm_first": false,
            "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": "fe876e8eaeb1b01ac4a21a337536fd164c10159bf1c5d049ea2715cca855718d",
          "node_id": "candidate_0ec0b671b6d720c69814"
        },
        {
          "id": "d2_plain_pre",
          "role": "diagnostic",
          "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": "487f99a81970de56cd83e5ed249732dbfa7860e0a5389abfec6823e543153207",
          "node_id": "candidate_53b641693beab96ad7cd"
        },
        {
          "id": "d2_plain_post",
          "role": "diagnostic",
          "model": {
            "type": "image_classifier",
            "mode": "cnn",
            "width": 24,
            "depth": 2,
            "activation": "gelu",
            "residual": false,
            "norm": "layer",
            "norm_first": false,
            "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": 150,
            "total_samples_seen": 4800,
            "gradient_accumulation": 1,
            "clip_norm": 1.0,
            "schedule": "constant",
            "precision": "fp32",
            "tf32": false,
            "deterministic_algorithms": true,
            "warmup_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": "1cc4fb9069b96900a53da98e555bc232d23fb2f3ce8a62596cb897b67dd82899",
          "node_id": "candidate_6cb23fbad1da49b99e05"
        }
      ]
    }
  ]
}