{
  "schema_version": 1,
  "git_commit": null,
  "checkpoint": "model.pt",
  "checkpoint_sha256": "5952ab8fb4184cffa7580b2035bc269531350d27e95cbdb4d26c9d5b241066c2",
  "benchmark": "calibration.npz",
  "benchmark_sha256": "e1f3b99cd5ce91dfe1e7de44995e1e38bbf1372766457025eca589625e89281e",
  "policy_config": {
    "max_nfe": 4096,
    "max_breadth": 128,
    "held_t": 0.4,
    "renoise": 1.0,
    "check_every": 16,
    "residual_window": 64,
    "residual_metric": "euclidean",
    "confirmation_checks": 8,
    "require_prediction_stability": true,
    "min_nfe": 64,
    "batch_size": 128,
    "seed": 0,
    "restart_seed_offset": 1000003,
    "flops_per_forward": 896039936
  },
  "residual": "trailing-window mean of per-puzzle euclidean distance between consecutive denoiser distributions, averaged over non-clue cells",
  "oracle_used_for_test_selection": false,
  "kind": "residual_stopping_calibration",
  "num_puzzles": 1024,
  "max_samples_override": null,
  "model_step": 175000,
  "labels_used_for": "development-set threshold calibration only",
  "selected": {
    "threshold": 0.4277484118938446,
    "accuracy": 0.9931640625,
    "solved": 1017,
    "mean_nfe": 327.265625,
    "accepted_fraction": 0.994140625,
    "fixed_accuracy": 0.9931640625,
    "max_accuracy_drop": 0.001,
    "selection_objective": "minimize mean NFE subject to accuracy >= fixed-depth accuracy - max_accuracy_drop; tie-break by higher accuracy then lower threshold"
  },
  "candidate_thresholds": 501,
  "calibration_bank": "calibration_bank.npz",
  "threshold_sweep": "threshold_sweep.json"
}