{
  "baseline": {
    "metrics": {
      "accuracy": 0.5644171779141104,
      "average_precision": 0.7310146473284146,
      "balanced_accuracy": 0.5565888554216868,
      "brier_score": 0.20382910064244883,
      "count": 326,
      "expected_calibration_error": 0.08678660959363686,
      "f1": 0.22826086956521738,
      "false_negative_rate": 0.86875,
      "false_positive_rate": 0.018072289156626505,
      "fn": 139,
      "fp": 3,
      "log_loss": 0.591045865107666,
      "macro_f1": 0.462421033073207,
      "negative_count": 166,
      "positive_count": 160,
      "precision": 0.875,
      "recall": 0.13125,
      "roc_auc": 0.752183734939759,
      "specificity": 0.9819277108433735,
      "threshold": 0.7638751511278986,
      "tn": 163,
      "tp": 21
    },
    "name": "acoustic_logistic",
    "predictions": {
      "bytes": 71129,
      "path": "artifacts/partial-baseline/validation_predictions.jsonl",
      "sha256": "76d67e2c35ac93d9898200f02d8626280b2139a623d432adf1266c9811980626"
    }
  },
  "bootstrap": {
    "samples": 2000,
    "seed": 17,
    "unit": "audit leakage group"
  },
  "candidate": {
    "metrics": {
      "accuracy": 0.5306748466257669,
      "average_precision": 0.6880689959324726,
      "balanced_accuracy": 0.5222138554216867,
      "brier_score": 0.2133037066638398,
      "count": 326,
      "expected_calibration_error": 0.1055241439925747,
      "f1": 0.11560693641618497,
      "false_negative_rate": 0.9375,
      "false_positive_rate": 0.018072289156626505,
      "fn": 150,
      "fp": 3,
      "log_loss": 0.6135001064370812,
      "macro_f1": 0.39809574378220525,
      "negative_count": 166,
      "positive_count": 160,
      "precision": 0.7692307692307693,
      "recall": 0.0625,
      "roc_auc": 0.7389683734939759,
      "specificity": 0.9819277108433735,
      "threshold": 0.741000771522522,
      "tn": 163,
      "tp": 10
    },
    "name": "tiny_tcn",
    "predictions": {
      "bytes": 106299,
      "path": "artifacts/partial-shard-warmstart-lr3e4-5ep/validation_predictions.jsonl",
      "sha256": "9eb4c01a9986efc169155c5e132afee9f2b1fe935c05809e10196ccdf6f70ab9"
    }
  },
  "candidate_minus_baseline_95ci": {
    "average_precision": {
      "bootstrap_fraction_delta_above_zero": 0.152,
      "estimate": -0.042945651395941975,
      "lower": -0.11945249084849802,
      "upper": 0.03523632760188885
    },
    "false_positive_rate": {
      "bootstrap_fraction_delta_above_zero": 0.425,
      "estimate": 0.0,
      "lower": -0.028409090909090908,
      "upper": 0.02976190476190476
    },
    "recall": {
      "bootstrap_fraction_delta_above_zero": 0.011,
      "estimate": -0.06875,
      "lower": -0.13297793599323277,
      "upper": -0.01219512195121951
    },
    "roc_auc": {
      "bootstrap_fraction_delta_above_zero": 0.3425,
      "estimate": -0.013215361445783058,
      "lower": -0.07823324414295993,
      "upper": 0.05120708236445977
    }
  },
  "count": 326,
  "fpr_budget": 0.02,
  "group_count": 326,
  "negative_count": 166,
  "positive_count": 160,
  "scope": "paired development comparison; not independent-test evidence",
  "threshold_note": "Each threshold was selected on this same development set. Bootstrap samples use those fixed thresholds; intervals do not remove model-selection or calibration bias."
}