{
  "schema_version": 1,
  "measured_at": "2026-10-19",
  "scope": "global_published_rule_sample",
  "status": "partial",
  "population_size": 6450,
  "sample_size": 385,
  "assessed": 101,
  "unassessed": 284,
  "correct": 85,
  "errors": 16,
  "error_breakdown": {
    "transcription": 2,
    "omission": 11,
    "applicability": 2,
    "stale": 0,
    "citation": 1
  },
  "source_agreement": 0.8415841584158416,
  "wilson_95": {
    "lower": 0.7580641570944454,
    "upper": 0.9000724310086231,
    "confidence": 0.95
  },
  "overall_accuracy": null,
  "sample_identification_bounds": {
    "lower": 0.22077922077922077,
    "upper": 0.9584415584415584
  },
  "cohen_kappa_rule_kind": {
    "value": 0.69630557294928,
    "n": 97,
    "observed_agreement": 0.8969072164948454,
    "chance_agreement": 0.6605377829737485,
    "reason": null
  },
  "unassessed_by_reason": {
    "source_unavailable": 234,
    "insufficient_source": 40,
    "superseded_source_applicability": 10,
    "pending_adjudication": 0
  },
  "measurement_description": "Agreement of sampled published rule statements with independently extracted source documents. Rule-kind kappa measures categorical answer type, not numerical accuracy. Unassessed rules prevent an overall population accuracy claim; conditional Wilson intervals do not correct nonresponse bias.",
  "method": "Stratified random allocation (multivariate hypergeometric), SRS without replacement within city/topic/district-category strata; equal inclusion probability n/N.",
  "blind_method": "Claude Opus 5.5 received only city, district, topic and primary-source text, without published values, citations or golden facts, in isolated sessions with tools disabled. A separate source comparison pass scored the frozen publication; a GPT-family adjudication reviewed all flagged errors. Historical drafting-model identity is unknown; independence is established by the blind prompt lineage and access isolation. Supported comparisons did not receive an additional native adjudication.",
  "source_vintage": "Source-document agreement only. Archived or codified documents do not establish absence of later adopted amendments. Gallatin previous-code selections are unassessed.",
  "limitations": [
    "Only the assessed subset contributes to the conditional source-agreement percentage and Wilson interval. Source availability is not random; the interval is not an overall accuracy interval.",
    "The original 385 selections remain the denominator for completeness. No replacement sampling and no post-correction score substitution.",
    "Cohen kappa concerns the categorical rule-answer type among resolved scored pairs, not numeric correctness or legal outcomes.",
    "Public-decision comparisons and National Zoning Atlas comparisons are separate from this random sample.",
    "All flagged errors were adjudicated; supported assessments did not receive a further independent adjudication. Automated comparison can still contain false negatives."
  ],
  "next_measurement_due": "2027-01-19",
  "report_url": "https://zoningverdict.com/how-we-verify#measured-accuracy",
  "evidence_url": "https://zoningverdict.com/accuracy/2026-10-19-audit.json",
  "fix_status": "Confirmed errors are tracked in the accompanying adjudication and correction evidence; this score remains the pre-correction baseline."
}
