{"version":"ew-v3","registeredAt":"2026-10-12","status":"Outcome-free feature registry; subsequent 2022–2025 analyses are exploratory","hypothesis":"Do public area features improve held-out police-reported motor-involved injury-collision frequency predictions beyond geography, road mix, traffic, density and exposure quality?","scope":"All 35,672 England and Wales LSOA2021 rows; membership independent of collisions, mileage or model eligibility.","outcome":"Distinct police-reported injury collisions with at least one in-scope motor vehicle, with log(all-motor vehicle miles / 100000) offset. Vehicle and collision records must not be counted interchangeably.","developmentYears":[2022,2023],"selectionYear":2024,"exploratoryAssessmentYear":2025,"excludedPrimaryYears":[2020,2021],"previousOutcomeAccess":"Existing social, parking and frost studies inspected 2023–2025 results. No 2025 claim of untouched confirmation is permitted.","freshReserve":"No fresh outcome reserve has been certified. Before prospective confirmation, audit prior access and register a still-uninspected period (candidate: final 2026), without altering this feature registry.","spatialValidation":"Leave one of nine English regions or Wales out; all areas of each UTLA remain together. Additional spatial dependence checks required before promotion.","models":["B: year, country/region, rurality, density, age composition, road class, terrain, traffic mix and exposure quality where supported","B + individual continuous feature","B + prespecified smooth continuous feature","B + categorical frozen deciles; original band retained if later model groups merge"],"fitting":"Poisson or negative-binomial counts with log-mileage offset; select dispersion and any regularisation using development data only. Same model-eligible rows and offsets in each comparison. Zero-collision positive-exposure rows retained.","missingness":"Do not remove master rows. Later modelling documents denominator-specific exclusions and missingness sensitivity. No collision-based feature eligibility.","proposedPromotionGate":"Research criterion fixed before new fits: >=1% held-out Poisson deviance reduction vs baseline, paired spatial-block 95% interval above zero, positive gain in each assessed year and >=7 of 10 geographic holdouts, overall and annual O/E between 0.90 and 1.10. Record unsupported bands and failures. Does not establish insurer transfer.","uncertainty":"Paired UTLA spatial-block bootstrap, 1000 replicates, seed 20261012; refit model and selection where relevant. Claims transfer requires a separate insurer test.","historicalAvailability":"This is a retrospective mixed-vintage research registry, not an as-of 2022 underwriting score. Census tables released in 2022/2023 and April 2026 roads must not be backdated. Historical as-of snapshots are a separate gate.","featureSelection":"Source-only E4–E7 construction. No risk-index weights learned. Post-2025 source vintages require retrospective or future-origin design; no temporal leakage in prospective testing.","exposureEndogeneity":"Flag Census, roads and rurality if they also enter exposure estimation; test independently measured traffic subsets and exposure sensitivity before claiming incremental accident risk.","excludedComposites":"Overall IMD and living-environment domain excluded from independent social evidence because of STATS19 outcome overlap; country deprivation ranks not pooled.","insurerBoundary":"Area of collision is not driver residence. No claim frequency, causal effect, fault, theft, fraud or commercial rating validation is claimed.","expansionFeatures":23,"newCalculatedFeatures":11}