{
  "schema_version": 1,
  "git_commit": null,
  "checkpoint": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/zebra_seed0_release_20261018_v1/bundles/zebra/model.pt",
  "checkpoint_sha256": "c43ce99d692dd8f0a15143bb28a35c69dd85f483ed430bca711ad6b865f7fe57",
  "benchmark": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/zebra_seed0_release_20261018_v1/bundles/zebra/test.npz",
  "benchmark_sha256": "0948626c109bafdff92f16c8ea3a2c8a6078c39960c1261fed28d3a7f7204431",
  "policy_config": {
    "max_nfe": 257,
    "max_breadth": 16,
    "held_t": 0.2,
    "renoise": 1.0,
    "check_every": 1,
    "residual_window": 16,
    "residual_metric": "euclidean",
    "confirmation_checks": 1,
    "require_prediction_stability": true,
    "min_nfe": 4,
    "batch_size": 64,
    "seed": 0,
    "restart_seed_offset": 1000003,
    "flops_per_forward": 0
  },
  "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_guided_inference_evaluation",
  "run_fingerprint": "ad3a2420cab1679dab9f04d8f963fc6acf0bfce272b1c233e0e0d84b10810112",
  "num_puzzles": 1024,
  "max_samples_override": null,
  "model_step": 115000,
  "calibration": "/proj/grp-flow-map/LoopedFlows/experiments/public_simplified_repro/outputs/zebra_seed0_release_20261018_v1/bundles/zebra/threshold.json",
  "calibration_sha256": "d7401816ce4e2638c587f5793f55b60a1110acb03a87941f71b1b05aa7bb46c3",
  "calibration_benchmark_sha256": "b2c66c97d53152de289a2ab0c4ad2123ec5bed4bf53aaab14f21fa7f90c527cd",
  "threshold": 0.10050372427701933,
  "threshold_selection": {
    "threshold": 0.10050372427701933,
    "accuracy": 1.0,
    "solved": 256,
    "mean_nfe": 4.51171875,
    "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"
  },
  "policies": [
    {
      "policy": "Fixed-depth FRM",
      "solved": 1022,
      "num_puzzles": 1024,
      "solve_rate": 0.998046875,
      "mean_nfe": 257.0,
      "mean_tflops": null
    },
    {
      "policy": "+ Residual early stopping",
      "solved": 1020,
      "num_puzzles": 1024,
      "solve_rate": 0.99609375,
      "mean_nfe": 6.240234375,
      "mean_tflops": null
    },
    {
      "policy": "+ Sequential restarts",
      "solved": 1020,
      "num_puzzles": 1024,
      "solve_rate": 0.99609375,
      "mean_nfe": 6.240234375,
      "mean_tflops": null
    }
  ],
  "accepted_by_threshold": 1024,
  "unresolved_at_breadth_cap": 0,
  "attempts": [
    {
      "attempt": 1,
      "attempted": 1024,
      "accepted": 1024,
      "remaining": 0,
      "mean_nfe_among_attempted": 6.240234375
    }
  ],
  "restart_frontier": [
    {
      "breadth_cap": 1,
      "solved": 1020,
      "num_puzzles": 1024,
      "solve_rate": 0.99609375,
      "mean_nfe": 6.240234375,
      "mean_tflops": null,
      "accepted": 1024,
      "unresolved": 0
    }
  ],
  "selection_rule": "stop each trajectory once the trailing-mean residual remains <= threshold for 1 checks and the decoded grid remains unchanged; restart only trajectories that reach max_nfe without acceptance; if max_breadth is exhausted, return the attempted terminal prediction with minimum residual",
  "labels_used_for": "final scoring only; never stopping, restart, or candidate selection"
}