{
  "schema_version": 1,
  "git_commit": null,
  "checkpoint": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/runs/maze_unique_20261016_v1/full/s0/stage_b/best.pt",
  "checkpoint_sha256": "b3115f2954f2d4d92daf7a893a80e1980fdb26f0bc813fb4944bd6f2abb8f698",
  "benchmark": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/benchmarks/maze_unique_20261016/calibration.npz",
  "benchmark_sha256": "3f0d8261ef2977c8a8d219dcb3f6cd146d3e7195438cfbe8939e774d1d8ff0dc",
  "policy_config": {
    "max_nfe": 257,
    "max_breadth": 16,
    "held_t": 0.3,
    "renoise": 1.0,
    "check_every": 1,
    "residual_window": 16,
    "residual_metric": "euclidean",
    "confirmation_checks": 2,
    "require_prediction_stability": true,
    "min_nfe": 4,
    "batch_size": 64,
    "seed": 0,
    "restart_seed_offset": 1000003,
    "flops_per_forward": 11336290304
  },
  "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": 256,
  "max_samples_override": null,
  "model_step": 30000,
  "labels_used_for": "development-set threshold calibration only",
  "selected": {
    "threshold": 0.010557100176811218,
    "accuracy": 1.0,
    "solved": 256,
    "mean_nfe": 5.0,
    "accepted_fraction": 1.0,
    "fixed_accuracy": 1.0,
    "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"
}