{
  "decision": "no-ship",
  "change": "lower semantic acceptance threshold from 0.72 to 0.62",
  "checks": {
    "candidate_coverage_at_least_20_percent": true,
    "candidate_strict_precision_at_least_80_percent": false,
    "precision_drop_no_more_than_5_points": false,
    "review_complete": true
  },
  "candidate_0_62": {
    "accepted": 35,
    "coverage": 0.23178807947019867,
    "abstention": 0.7682119205298014,
    "strict_correct": 24,
    "strict_precision": 0.6857142857142857,
    "strict_precision_wilson_95": [
      0.5202023638793081,
      0.8144915532573418
    ],
    "correct_or_partial": 30,
    "concept_fit_precision": 0.8571428571428571
  },
  "conservative_0_72": {
    "accepted": 18,
    "coverage": 0.11920529801324503,
    "abstention": 0.880794701986755,
    "strict_correct": 14,
    "strict_precision": 0.7777777777777778,
    "strict_precision_wilson_95": [
      0.5478541568378739,
      0.9099907189139831
    ],
    "correct_or_partial": 16,
    "concept_fit_precision": 0.8888888888888888
  },
  "exact_token_baseline": {
    "accepted": 9,
    "coverage": 0.059602649006622516,
    "abstention": 0.9403973509933775,
    "strict_correct": 5,
    "strict_precision": 0.5555555555555556,
    "strict_precision_wilson_95": [
      0.26665129349549305,
      0.8112214789023355
    ],
    "correct_or_partial": 7,
    "concept_fit_precision": 0.7777777777777778
  },
  "candidate_minus_conservative_strict_precision": -0.09206349206349207,
  "limitations": [
    "Human adjudication is new, but the holdout and model judgments previously informed development.",
    "The evaluation measures prediction quality on repository-disjoint components, not end-user outcomes."
  ]
}