{
  "version": "ew-phase3-v1",
  "frozenBeforeNewOutcomeJoin": "2026-10-12",
  "phase": 3,
  "years": [
    2022,
    2023,
    2024,
    2025
  ],
  "masterCount": 35672,
  "expectedRows": 142688,
  "featureRelease": "ew-v4",
  "featureRankings": "Reused unchanged; never rerank subset or outcome sample.",
  "numerator": "Distinct police-reported injury collisions with at least one definite in-scope motor vehicle; unresolved motor states are sensitivity bounds. Unique full LSOA2021 polygon covers assignment; shared boundaries unassigned.",
  "denominator": "All-motor vehicle miles on access-screened OS April 2026 roads. Annual major AADF \u00d7 published km /1.609344 \u00d7 days; independent minor residual allocation reconciled to official authority totals.",
  "majorRouteGates": {
    "sameRoadNumber": true,
    "endpointSnapMetres": 100,
    "countPointDistanceMetres": 75,
    "routeLengthRatio": [
      0.8,
      1.2
    ],
    "ownAuthorityShare": 0.99
  },
  "majorFallback": "Same number within control area, then same major class; retained and flagged, never described as routed.",
  "controls": "UTLA23 master. In 2022 the 319 Cumberland/Westmorland LSOAs share the published former Cumbria control total; no invented successor totals.",
  "minorModel": {
    "developmentYears": [
      2022,
      2023
    ],
    "geographicFold": "SHA256 of 20km BNG block, first eight hex digits modulo five; canonical site block",
    "candidate": "Regularised Poisson mean alpha=1, site inverse-frequency weights",
    "features": [
      "region",
      "roadClass",
      "function",
      "logDensity",
      "binaryLSOAUrbanIndicator",
      "meanEndpointDegree"
    ],
    "benchmark": "Region \u00d7 class \u00d7 resident settlement means shrunk by 20 sites to class means",
    "selection": "Development geographic OOF Poisson deviance only; prefer benchmark within 1% of fitted model",
    "futureDiagnostics": [
      2024,
      2025
    ],
    "alternatives": [
      "fitted flow \u00d7 road length",
      "shrunk benchmark flow \u00d7 road length",
      "road length only (no resident covariates)"
    ]
  },
  "quality": {
    "strictMajorFallbackMaximum": 0.05,
    "minimumStratumSites": 30,
    "minimumValidationBlocks": 3,
    "trafficBiasGate": 0.1,
    "validatedFrequencyEligible": false
  },
  "vintages": "2026 road/access and census/urban resident context reused retrospectively; not historic road openings, closures, restrictions or road-specific urban classification.",
  "limitations": [
    "Control-total agreement does not validate LSOA allocation.",
    "Minor residual absorbs major reconstruction discrepancies and is not an independently measured minor total.",
    "Density/urbanity/road function enter exposure; later tests of related features require denominator sensitivity.",
    "All features have mixed vintages; source-vintage register remains mandatory.",
    "2025 has already been examined and is exploratory, not untouched validation.",
    "No accident or claims model is fitted in Phase 3.",
    "Police injury collision rates are not all insurance claim rates."
  ]
}