schema_version: vision_question_clusters_v1
description: One bounded supplementary image-generation cluster selected using original discovery validation only.
  Independent new data/training seeds and multiplicity family; original rejected triples remain unchanged.
dataset_family: conditional_image_generation
selection_metric: test_energy_score
base_profile: image-generation-ranking-base-v1.yaml
default_num_choices: 3
allowed_num_choices:
- 3
- 4
- 5
instance_seeds:
- 852001
- 852002
- 852003
discovery_training_seeds:
- 40
- 41
- 42
- 43
confirmation_training_seeds:
- 240
- 241
- 242
- 243
- 244
- 245
- 246
- 247
- 248
- 249
- 250
- 251
- 252
- 253
- 254
- 255
statistics:
  discovery:
    n_seeds: 4
    split: validation
    method: direction_estimation
    ordering: mean_validation_energy_score
    admission_gate: false
    purpose: Freeze adjacent ranking directions before independent confirmation; no significance claim from four
      discovery seeds.
  confirmation:
    n_seeds: 16
    split: test
    test: one_sided_paired_sign_test
    alpha: 0.05
    correction: holm
    minimum_mean_gap: 0.003
  estimand: Training-seed paired direction stability conditional on one fixed dataset and fixed evaluation-noise/reference
    stream, with observed mean energy-score gap; not a test of expected-score mean difference or a natural-image
    population claim.
selection_contract:
  discovery_split: validation
  confirmation_split: test
  minimum_mean_gap: 0.003
  require_nonoverlap: false
  multiplicity: 'Independent supplement family: Holm across all six frozen adjacent hypotheses, including rejected
    instances.'
  claim_scope: Fixed synthetic image distribution and fixed independent evaluation sample streams; training-seed
    robustness only.
  difficulty_policy: Evaluate every admitted supplemental triple and report all rejects separately from the original
    fixed batch; no answer-based selection or replacement.
defaults:
  dataset:
    size: 16
    noise_steps: 32
    condition_dim: 6
    brightness_min: 0.65
    brightness_max: 1.0
    texture_strength: 0.12
    radius_min: 0.25
    radius_max: 0.42
    position_extent: 0.45
    reference_samples: 4
    train_size: 1024
    validation_size: 64
    test_size: 64
    ood_size: 32
    evaluation_shift: {}
    ood_shift:
      position_extent: 0.55
  optimizer:
    type: AdamW
    lr: 0.001
    weight_decay: 0.01
    betas:
    - 0.9
    - 0.999
    eps: 1.0e-08
    foreach: false
  loss:
    loss_id: diffusion_mse
  training:
    batch_size: 16
    training_steps: 3000
    total_samples_seen: 48000
    gradient_accumulation: 1
    clip_norm: 1.0
    schedule: constant
    precision: fp32
    tf32: false
    deterministic_algorithms: true
    warmup_updates: 0
    batch_rng_device: cpu
    paired_cpu_batch_stream: true
    persist_initial_state: true
    batch_seed_offset: 500000
    noise_seed_offset: 600000
    evaluation_split: validation
    eval_steps:
    - 750
    - 1500
    - 3000
    prediction_type: x0
    noise_steps: 32
    noise_schedule: cosine
    loss_weighting: uniform
    min_snr_gamma: 5.0
    sampling_steps: 16
    sampling_eta: 0.0
    evaluation_samples_per_condition: 4
    evaluation_seed: 314159
    evaluation_batch_conditions: 16
models:
  conv:
    type: image_generation_denoiser
    mode: conv
    width: 24
    depth: 3
    kernel: 3
    norm: group
    activation: silu
    residual: true
    block: dense
    conditioning: concat
    coordinates: absolute
    time_embedding: sinusoidal
    max_parameters: 2000000
  unet:
    type: image_generation_denoiser
    mode: unet
    width: 24
    depth: 3
    kernel: 3
    norm: group
    activation: silu
    residual: true
    block: dense
    skip: true
    conditioning: concat
    coordinates: absolute
    time_embedding: sinusoidal
    max_parameters: 2000000
  mlp:
    type: image_generation_denoiser
    mode: mlp
    width: 24
    depth: 3
    activation: silu
    conditioning: concat
    coordinates: absolute
    time_embedding: sinusoidal
    max_parameters: 2000000
  transformer:
    type: image_generation_denoiser
    mode: patch_transformer
    width: 24
    depth: 3
    activation: gelu
    conditioning: concat
    coordinates: absolute
    time_embedding: sinusoidal
    patch_size: 2
    max_parameters: 2000000
    num_heads: 4
    ff_multiplier: 2
contexts:
  textured: {}
  clean:
    texture_strength: 0.0
  small_data:
    train_size: 96
  shifted:
    evaluation_shift:
      position_extent: 0.55
surface_heuristics:
- id: more_parameters
  field: parameter_count
  higher_is_better: true
- id: more_training_steps
  field: training.training_steps
  higher_is_better: true
- id: more_sampling_steps
  field: training.sampling_steps
  higher_is_better: true
templates:
- id: operator_native_matched
  category: operator
  context_ref: textured
  budget_semantics: matched_steps
  hypothesis: Four native denoisers share absolute coordinate input, data, optimizer, target and update budget.
    Their operator, parameter count, normalization and connectivity differ as shown; finite-budget ordering requires
    actual sampled image quality.
  varying_axes:
  - model.mode
  choices:
  - id: conv
    model_ref: conv
  - id: unet
    model_ref: unet
  - id: mlp
    model_ref: mlp
  diagnostics:
  - id: transformer
    model_ref: transformer
  controls:
  - 'All four available native operator cells are retained: convolution, one-level U-Net, flat MLP and patch Transformer.'
  - Every cell uses absolute coordinates, concat conditions, sinusoidal time, x0 target, 3000 updates, batch16,
    AdamW1e-3, the same materialized data and fixed evaluation noise/reference draws.
  - 'This is a bundled native-recipe contrast: parameter count, activation, normalization and connectivity are explicitly
    shown; do not attribute the total difference exclusively to operator name.'
  - Primary choices were selected from original discovery validation only; the diagnostic Transformer is not promoted
    or substituted after new confirmation.
  shortcut_checks:
  - always prefer U-Net for diffusion
  - always prefer more parameters
  - always prefer globally mixing operator
  explanation_target: Explain both adjacent comparisons using exact native code and the fixed finite-budget task;
    use all four local cells and image/geometry/diversity diagnostics, while acknowledging unresolved parameter-count
    and optimization attribution.
  admission: Freeze the new four-seed validation order, then independently test all six adjacent hypotheses from
    three new dataset instances using the unchanged16-seed sign-Holm rule and gap0.003. Keep rejected instances;
    no replacement.
selection_provenance:
  source_dsl: profiles/image-generation-ranking-clusters-v1.yaml
  source_dsl_sha256: 9066b70f1272f930381c8c017b7daf9d74ce93bc79d4eb6204395e08c0a0470e
  source_plan_sha256: 004fa55685f56051cb0de5788733a3c476ca2f8721d9c1ee6ec7d7c0b55403eb
  selection_policy: data/task_expansion/vision_batch/image_ranking_v1_20261015/generation/supplement_proposal/selection_policy.json
  selection_policy_sha256: b9d104c2474895543a9159e723e9112d2d6a19214fe940d763a6492f963b75f4
  selection_evidence: data/task_expansion/vision_batch/image_ranking_v1_20261015/generation/supplement_proposal/selection_evidence.json
  selection_evidence_sha256: 01768104dc3c38a11175697129407e1c5a4b853bac0dbfa7b76560396aafcf79
  selector_sha256: 661858113970e129c33d9c91889a8d7d7c8dd7ceba953de77fd188916ee30cef
  only_metric: original discovery final_validation_metric
  enumerated_triples: 33
  unchanged_original_triples_excluded: 4
  selected_new_clusters: 1
  selection_reads_confirmation_outputs: false
  selection_reads_llm_answers: false
  scope_limit: The analyst already knew original batch aggregate admission and quality; candidate selection is independently
    replayable using discovery validation alone. This is a validation-selected supplementary batch, not a repair
    or replacement of the original frozen test.