schema_version: vision_question_clusters_v1
base_profile: task-expansion-image-classification.yaml
dataset_family: image_classification
selection_metric: test_cross_entropy
default_num_choices: 3
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.
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.
defaults:
  dataset:
    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
  optimizer:
    type: AdamW
    lr: 0.001
    weight_decay: 0.01
    betas:
    - 0.9
    - 0.999
    eps: 1.0e-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:
    - 0
    - 100
    - 300
    record_training_diagnostics: true
models:
  cnn:
    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
  vit:
    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
  mlp:
    type: image_classifier
    mode: mlp
    width: 96
    depth: 2
    activation: gelu
    residual: true
    max_parameters: 2000000
description: 'Independent bounded supplement: four predeclared image classification clusters selected
  from v1 validation evidence, without reading model answers. All new data and training seeds; no reuse
  of v1 significance. Analyst-only hypotheses never enter prompts.'
allowed_num_choices:
- 3
instance_seeds:
- 842901
- 842902
- 842903
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
contexts:
  relations: {}
  shapes:
    task: local_shape
provenance:
  source_of_design: v1 validation trajectories and complete local training controls only; no model-answer-conditioned
    selection
  no_llm_responses_read: true
  source_dsl_sha256: a9f978cb00d7650c29c7389e5fce4b6945945e727c66ce2ca22de6e2b9a2c627
  validation_evidence_path: data/task_expansion/vision_validation/classification_bounded_supplement_validation_evidence.json
  validation_evidence_sha256: c1d3f6848ec3c6865754466c8b2469e6ae04863da0ecbfe477dab48b57a6801c
  independent_batch: true
  reuse_v1_significance: false
  frozen_template_limit: 4
  frozen_instance_count: 12
  frozen_adjacent_hypothesis_count: 24
  stop_rule: Evaluate all admitted triples after one fixed 4+16-seed run; retain rejections without replacements
    or tuning.
templates:
- id: operator_learning_speed
  category: operator_budget
  context_ref: relations
  budget_semantics: matched_steps
  varying_axes:
  - model.mode
  choices:
  - id: mlp_75
    model_ref: mlp
    training_updates:
      training_steps: 75
  - id: cnn_75
    model_ref: cnn
    training_updates:
      training_steps: 75
  - id: vit_75
    model_ref: vit
    training_updates:
      training_steps: 75
  diagnostics:
  - id: mlp_300
    model_ref: mlp
    training_updates:
      training_steps: 300
  - id: cnn_300
    model_ref: cnn
    training_updates:
      training_steps: 300
  - id: vit_300
    model_ref: vit
    training_updates:
      training_steps: 300
  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.
- id: operator_shape_generalization
  category: operator_generalization
  context_ref: shapes
  budget_semantics: matched_steps
  varying_axes:
  - model.mode
  - optimizer.lr
  choices:
  - id: cnn_spatial_lr003
    model_ref: cnn
    training_updates:
      training_steps: 1200
    optimizer_updates:
      lr: 0.003
  - id: vit_lr001
    model_ref: vit
    training_updates:
      training_steps: 1200
  - id: mlp_lr001
    model_ref: mlp
    training_updates:
      training_steps: 1200
  diagnostics:
  - id: cnn_spatial_lr001
    model_ref: cnn
    training_updates:
      training_steps: 1200
  - id: cnn_mean_lr001
    model_ref: cnn
    training_updates:
      training_steps: 1200
    model_updates:
      pooling: mean
  - id: cnn_mean_lr003
    model_ref: cnn
    training_updates:
      training_steps: 1200
    model_updates:
      pooling: mean
    optimizer_updates:
      lr: 0.003
  - id: vit_lr003
    model_ref: vit
    training_updates:
      training_steps: 1200
    optimizer_updates:
      lr: 0.003
  - id: mlp_lr003
    model_ref: mlp
    training_updates:
      training_steps: 1200
    optimizer_updates:
      lr: 0.003
  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.
- id: patch_coordinate_retention
  category: patch_position
  context_ref: relations
  budget_semantics: matched_steps
  varying_axes:
  - model.patch_size
  choices:
  - id: p3_none
    model_ref: vit
    training_updates:
      training_steps: 300
    model_updates:
      patch_size: 3
      position: none
  - id: p6_none
    model_ref: vit
    training_updates:
      training_steps: 300
    model_updates:
      patch_size: 6
      position: none
  - id: p12_none
    model_ref: vit
    training_updates:
      training_steps: 300
    model_updates:
      patch_size: 12
      position: none
  diagnostics:
  - id: p3_learned
    model_ref: vit
    training_updates:
      training_steps: 300
    model_updates:
      patch_size: 3
  - id: p6_learned
    model_ref: vit
    training_updates:
      training_steps: 300
    model_updates:
      patch_size: 6
  - id: p12_learned
    model_ref: vit
    training_updates:
      training_steps: 300
    model_updates:
      patch_size: 12
  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.
- id: residual_norm_placement
  category: residual_normalization
  context_ref: relations
  budget_semantics: matched_steps
  varying_axes:
  - model.residual
  - model.norm_first
  choices:
  - id: d6_residual_pre
    model_ref: cnn
    training_updates:
      training_steps: 150
    model_updates:
      depth: 6
  - id: d6_residual_post
    model_ref: cnn
    training_updates:
      training_steps: 150
    model_updates:
      depth: 6
      norm_first: false
  - id: d6_plain_pre
    model_ref: cnn
    training_updates:
      training_steps: 150
    model_updates:
      depth: 6
      residual: false
  diagnostics:
  - id: d6_plain_post
    model_ref: cnn
    training_updates:
      training_steps: 150
    model_updates:
      depth: 6
      residual: false
      norm_first: false
  - id: d2_residual_pre
    model_ref: cnn
    training_updates:
      training_steps: 150
  - id: d2_residual_post
    model_ref: cnn
    training_updates:
      training_steps: 150
    model_updates:
      norm_first: false
  - id: d2_plain_pre
    model_ref: cnn
    training_updates:
      training_steps: 150
    model_updates:
      residual: false
  - id: d2_plain_post
    model_ref: cnn
    training_updates:
      training_steps: 150
    model_updates:
      residual: false
      norm_first: false
  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.