{
  "baseline": {
    "n": 90,
    "correct": 85,
    "accuracy": 0.9444444444444444,
    "accuracy_ci95": [
      0.8764626377820072,
      0.9760390283830375
    ],
    "p50_ms": 97.68479151534848,
    "p95_ms": 150.9270685928641,
    "calibration": {
      "temperature": 2.046065578288829,
      "ece": 0.05596926803442395,
      "mce": 0.33236392037060847,
      "brier_score": 0.07830253878305025,
      "nll": 0.1677882594430572,
      "accuracy": 0.9444444444444444,
      "passes_gate": false,
      "bin_accuracies": [
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.4,
        1.0,
        1.0,
        0.8,
        1.0
      ],
      "bin_confidences": [
        0.05,
        0.15000000000000002,
        0.25,
        0.35000000000000003,
        0.45,
        0.538938836816857,
        0.6676360796293915,
        0.7744787175327771,
        0.8549314118822144,
        0.9705751240130236
      ],
      "bin_counts": [
        0,
        0,
        0,
        0,
        0,
        5,
        2,
        5,
        10,
        68
      ]
    }
  },
  "cyclic": {
    "n": 90,
    "correct": 89,
    "accuracy": 0.9888888888888889,
    "accuracy_ci95": [
      0.9397159168824676,
      0.9980359158990815
    ],
    "p50_ms": 291.43383298651315,
    "p95_ms": 453.5067100514425,
    "calibration": {
      "temperature": 0.1000059200037425,
      "ece": 0.01599944151227259,
      "mce": 0.44810728267265465,
      "brier_score": 0.026660513062642943,
      "nll": 0.09002366421751616,
      "accuracy": 0.9888888888888889,
      "passes_gate": true,
      "bin_accuracies": [
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        1.0,
        0.0,
        0.0,
        0.0,
        0.9887640449438202
      ],
      "bin_confidences": [
        0.05,
        0.15000000000000002,
        0.25,
        0.35000000000000003,
        0.45,
        0.5518927173273454,
        0.6500000000000001,
        0.75,
        0.8500000000000001,
        0.9999083421733919
      ],
      "bin_counts": [
        0,
        0,
        0,
        0,
        0,
        1,
        0,
        0,
        0,
        89
      ]
    }
  },
  "tournament": {
    "n": 90,
    "correct": 87,
    "accuracy": 0.9666666666666667,
    "accuracy_ci95": [
      0.9065271396961037,
      0.9885996097771934
    ],
    "p50_ms": 485.4218955151737,
    "p95_ms": 747.5381311145611,
    "calibration": {
      "temperature": 0.8140395946537136,
      "ece": 0.02571698233240563,
      "mce": 0.610999927290451,
      "brier_score": 0.03845998955026209,
      "nll": 0.06777357373448877,
      "accuracy": 0.9666666666666667,
      "passes_gate": true,
      "bin_accuracies": [
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        0.0,
        1.0,
        0.8,
        1.0
      ],
      "bin_confidences": [
        0.05,
        0.15000000000000002,
        0.25,
        0.35000000000000003,
        0.40610426431391683,
        0.55,
        0.610999927290451,
        0.7143802177079246,
        0.8710684623216698,
        0.9919943643364425
      ],
      "bin_counts": [
        0,
        0,
        0,
        0,
        1,
        0,
        1,
        1,
        5,
        82
      ]
    }
  },
  "autonomous": {
    "n": 90,
    "correct": 87,
    "accuracy": 0.9666666666666667,
    "accuracy_ci95": [
      0.9065271396961037,
      0.9885996097771934
    ],
    "coverage": 1.0,
    "threshold": 0.0,
    "calibration_accepted": 30,
    "policy": "Largest calibration coverage with >=92% empirical accuracy and >=10 examples; freeze threshold before evaluation. If none qualifies, abstain all."
  },
  "combined": {
    "n": 90,
    "correct": 87,
    "accuracy": 0.9666666666666667,
    "accuracy_ci95": [
      0.9065271396961037,
      0.9885996097771934
    ],
    "committed": 90,
    "abstentions": 0,
    "coverage": 1.0,
    "committed_accuracy": 0.9666666666666667,
    "p50_ms": 485.42220797389746,
    "p95_ms": 747.5383580342168,
    "ece": null,
    "calibration_note": "Bayesian assumed posteriors not reported as calibrated confidence; combined ECE not measured.",
    "latency_note": "Measured tournament latency plus measured subsequent routing/probe time; components run in separate evaluation passes, not a single deployed endpoint."
  },
  "order_effects": {
    "n": 90,
    "raw_rotation_disagreements": 10,
    "raw_rotation_disagreement_rate": 0.1111111111111111,
    "reversed_orbit_disagreements": 1,
    "reversed_orbit_disagreement_rate": 0.011111111111111112,
    "mean_max_probability_difference": 0.016417959894842453,
    "worst_max_probability_difference": 0.9839703740898819,
    "definition": "Raw canonical winner disagreement across three rotations; residual difference between opposite three-way cyclic orbits. Cyclic invariance alone is algebraic, not a measurement of no bias."
  },
  "hashes": {
    "cases.jsonl": "2f7994d4be5d3b80a040dde3422d1dd1dda2422536d3dd79409598769dbe87f9",
    "predictions.jsonl": "9822b0abc5896e2c42d4d7990bd7c0849a509abdeaae54802dd0a0db2ae8cc49",
    "runtime.json": "305e2e8eb721053c8cb5d36c2f727e09625e53d1366aee0e9fafb5e59a690b20",
    "combined_decisions.json": "d672f1c0f0ea393dbcdb10a20c09b8934670ae5690969398cbecbb22d084f9be",
    "always_probe_decisions.json": "c2716c74174905243e95bb67a52763fd2278f03b96368ee02aea0a1cd31e7d1f"
  },
  "split_hashes": {
    "calibration": "ed88360744067636abcc947234fc9126e749dd607b7e5e820bc0b441f5004c39",
    "evaluation": "e2f22cae0590df7c2473d24bb4dd4b030489b186f3a19d713aca2489a0768d14"
  },
  "always_probe": {
    "n": 90,
    "correct": 87,
    "accuracy": 0.9666666666666667,
    "accuracy_ci95": [
      0.9065271396961037,
      0.9885996097771934
    ],
    "p50_ms": 570.9782080375589,
    "p95_ms": 829.8614019091474,
    "probe_p50_ms": 82.36499998020008,
    "scope": "Secondary always-probe diagnostic added after primary results; same cases/logits, first executed assertion on all candidates, separate measured sandbox stage. No threshold retuning.",
    "test_passing_candidates": 163
  }
}