{
  "version": "ew-phase4-v1",
  "registeredAt": "2026-10-13",
  "status": "Retrospective development; no independently validated index",
  "inputCommit": "c7bd306d9f6eef285354f9d6d7313fcb5b051da9",
  "outcome": "Distinct police-reported motor-involved injury collisions; no vehicle or casualty counts",
  "offset": "log(annual scenario all-motor vehicle miles / 100000)",
  "nationalGeography": 35672,
  "years": [
    2022,
    2023,
    2024,
    2025
  ],
  "featureDeciles": "All national feature values/bands remain frozen; no decile feature screening in this phase",
  "eligibility": "Phase3 analysisEligible AND all three finite positive denominators AND all supported control values; one common cohort for every candidate and scenario. Excluded rows retained in eligibility.csv.gz.",
  "development": [
    2022,
    2023
  ],
  "selectionDiagnosticYear": 2024,
  "exploratoryAssessmentYear": 2025,
  "selectionRule": "On fitted exposure only, average the ratios of candidate to core_poisson Poisson deviance for 2022\u21922023 and pooled leave-one-region-out 2022\u201323. Select the simplest model within 1% of the best score, complexity order core_poisson, core_nb2, rich_poisson, rich_nb2. Do not reselect using 2024/25 or alternative denominators.",
  "models": [
    "core_poisson",
    "core_nb2",
    "rich_poisson",
    "rich_nb2"
  ],
  "core": [
    "year factor",
    "sum-zero region effects (nine English regions and Wales)",
    "Wales \u00d7 year where estimable",
    "harmonised binary resident urban context",
    "log(1+residents/km\u00b2)",
    "Census age shares 0\u201314,15\u201324,65+",
    "static motorway and A-road length shares; minor is reference",
    "major fallback share",
    "unsupported/sparse/extrapolated minor share",
    "major estimated share of preferred total miles"
  ],
  "rich": [
    "core controls",
    "linear log-density hinges at training 10th,50th,90th percentiles; linear tails",
    "urban \u00d7 log density",
    "log(1+junctions/km)",
    "log(1+roundabout nodes/km)",
    "log(road sinuosity)"
  ],
  "unsupportedControls": [
    "road gradient/terrain",
    "HGV/van/motorcycle/bus mileage shares"
  ],
  "missingControlPolicy": "No substitute based on collision attributes. Supported controls complete; withheld feature candidates stay deferred.",
  "spatialValidation": "Leave each of nine English regions and Wales out, retaining all years and whole UTLAs. Fit only 2022\u201323 outside held region. Unseen region receives mean fitted log region effect (zero sum contrast); unestimable Wales interactions zero. This is a transportability stress test, not normal in-region performance.",
  "temporalValidation": "Fit 2022, predict 2023. Freeze selected 2022\u201323 fit for 2024 and 2025; latest fitted year effect and Wales-year effect carried forward. No test-year calibration fitted. Rolling 2025 benchmark separately allows earlier 2024 outcomes.",
  "countryPolicy": "Wales-year interactions absorb country-specific changes only where estimable. September 2023 default 20mph is context, not a causal covariate estimate. Carrying 2023 to later years does not represent a full-year policy effect.",
  "dispersion": "Compare Poisson and NB2 (Var=mu+alpha*mu\u00b2); NB alpha estimated by development maximum likelihood separately inside each fold. No penalty or alpha selected on future outcomes.",
  "priorBenchmark": "Separate posterior-style local multiplier (sum earlier observed +10)/(sum earlier baseline expected +10), using only earlier eligible outcomes. Fixed pseudo-count10. For strict spatial transportability no held-region history is used. For history-assisted spatial benchmark, held-region 2022 history may predict 2023 and is labelled explicitly.",
  "uncertainty": "Final-fit coefficient covariance clustered by UTLA with Cumbria successors joined permanently. Diagnostic O/E and paired deviance intervals resample these blocks 1000 times, seed20261012, conditional on fixed fitted predictions. No full fit/selection bootstrap and no promotion claim; E8 promotion requires refit uncertainty in a later phase.",
  "spatialDependence": "Moran I separately by year and on mean LSOA spatial-OOF Pearson residuals, directed8nearest centroid neighbours, row-standardised weights;199 permutations. Cross-authority version also reported. This diagnoses dependence; does not eliminate it.",
  "exposureSensitivity": "Refit selected structure/family separately on each of three denominators using identical rows/splits; report performance, O/E, fitted rate rank and coefficient changes. Uncertainty conditional on allocated exposure.",
  "roadScopeDifference": "Frozen motorway/A length shares include restricted-access roads; Phase3 allocated exposure excludes them. Rich road context is a retrospective April2026 proxy, not an aligned historical speed/road inventory.",
  "predictorWhitelist": "No collision-derived fields, qualityFlags string, road-match counts, boundary-collision flags or eligibility result enters X.",
  "measuredScopeSensitivity": "Separately fit counted accepted non-overlapping major-route cells with collisions matched to the same route and polygon fragments; annual AADF\u00d7physical routed length denominator. Excludes minor-road allocation, but remains an imperfect selected-route sensitivity.",
  "betterObservedSubset": "Outcome-free flag: counted major miles >=50% of preferred miles, major fallback <=5%, no shared major route sections. Evaluate matching all-road numerator/denominator; this is NOT an independently measured all-road subset and strictExposureEligible remains zero.",
  "calibrationGate": "O/E0.90\u20131.10, inherited diagnostic tolerance. Report Wales, each region, urban/rural, country\u00d7year, exposure-quality/support/spread and prediction-rate groups. Material subgroup failures restrict predictive claims even if aggregate passes.",
  "historicalVintage": "Census2021, April2026 roads and latest revised collision/mileage records; retrospective research, no as-of underwriting availability certified.",
  "sources": [
    {
      "url": "https://www.gov.wales/introducing-default-20mph-speed-limits",
      "use": "Welsh policy timing context only"
    },
    {
      "url": "https://www.statsmodels.org/stable/generated/statsmodels.discrete.discrete_model.NegativeBinomial.html",
      "use": "NB2 likelihood and log exposure offset"
    }
  ],
  "insurerBoundary": "Area-of-collision injury-frequency development baseline; not residence risk, not causality, not all-claims validation."
}