{
  "version": "ew-phase5-v1",
  "registeredDate": "2026-10-13",
  "sourceCommit": "752c7b53d10135ada28f150a98d1e6519ab53af8",
  "nationalLSOAs": 35672,
  "bandSource": "Frozen ew-v4 registry / exact ew-v3 full-feature lookup. No re-ranking, cutoffs or band merging in Phase5. Higher band means higher feature, not higher risk.",
  "developmentYears": [
    2022,
    2023
  ],
  "laterYears": "2024 selection-year diagnostic and 2025 exploratory only; neither can select representation or disposition. Both previously inspected; no untouched validation.",
  "outcome": "Distinct motor-involved police-recorded injury collisions",
  "baseline": "Frozen Phase4 rich_poisson structure, refitted on each feature-eligible cohort within every fold. Count offset log(miles/100000); all eligible zero counts retained.",
  "novelPrimaryFamily": [
    "no_qualifications",
    "level4_qualifications",
    "social_rented",
    "private_rented",
    "owner_occupied",
    "overcrowded_bedrooms",
    "unemployment",
    "terraced",
    "flats",
    "no_car",
    "cars_per_household_lower_bound",
    "air_frost_normal",
    "multi_car",
    "three_plus_car",
    "car_commute",
    "long_commute",
    "work_from_home",
    "walk_commute",
    "cycle_commute",
    "connectivity_public_transport",
    "connectivity_walking",
    "connectivity_cycling",
    "connectivity_driving",
    "air_no2",
    "air_pm25",
    "bankruptcy_la",
    "insolvency_la",
    "heat_summer_days_gt25"
  ],
  "baselineFormFamily": [
    "population_density",
    "age15_24",
    "age65_plus",
    "age0_14",
    "motorway_share",
    "a_road_share",
    "minor_road_share",
    "junction_density",
    "roundabout_density",
    "road_sinuosity"
  ],
  "forms": [
    "linear",
    "smooth",
    "categorical"
  ],
  "primaryInference": "One final-development categorical omnibus cluster-Wald p per feature. Benjamini\u2013Hochberg FDR separately across 28 novel and 10 baseline form diagnostics. Failed/unestimable p=1. No minimum p across forms.",
  "representation": "Primary screening and retention use categorical only. Report best development mean relative temporal/spatial deviance representation as secondary research evidence; no later-year selection or inferential p-value switching.",
  "continuous": "Linear and restricted cubic spline of the value; five training quantiles .05,.275,.5,.725,.95, 4 df with linear tails. Density/junctions/roundabouts log1p, sinuosity log; others identity. Sparse features add zero/positive indicator and conditional positive basis. Training transformations refitted inside folds.",
  "rank": "Keep every baseline column; residualise added bases against baseline and QR only the new columns. Zero incremental df is already represented.",
  "spatial": "10 leave-one-region-out folds in 2022\u201323: nine English regions and Wales. Unseen region has mean log-region effect as in Phase4. All UTLAs stay together.",
  "temporal": "Fit2022\u2192predict2023, then fit2022\u201323 for fixed 2024/25 exploratory diagnostics. Latest baseline time effects carried forward.",
  "cluster": "Union UTLAs sharing actual E7 source reporting area, plus Cumbria successors. Same permanent mapping for every test. Financial cards show independent source reporting units per band.",
  "uncertainty": "Final and subgroup coefficients: finite-sample corrected cluster sandwich. Paired deviance intervals: 1000 resamples of fixed prediction block totals, seed20261012. Excludes fit/selection and exposure uncertainty; explicitly does NOT meet E8 refit-bootstrap promotion requirement.",
  "support": "Flag bands with <100 LSOA-years OR <100 collisions OR <10 independent clusters. Empty national bands stay visible. If a held-out categorical band was absent in training, use baseline prediction, count/flag unsupported rows and block retention. This fallback is not an estimated neutral effect.",
  "exposure": "Same feature-eligible cohort for preferred, traffic-group benchmark and length offsets. Categorical temporal and final development models independently refitted for each. Also refit categorical on Phase4 counted-route scope, with matched numerator, physical counted mileage and scope-specific baseline road shares; screen feature itself remains frozen. Counted temporal check is not representative all-road validation. Better-observed all-road temporal subset also reported.",
  "persistence": "Separately refit categorical models in each development year, region and urban/rural context; region/urban contrasts exploratory, no independent significance claims. Also report held-out band residual patterns by year and subgroup.",
  "researchRetention": "Novel categorical q<=.05 AND mean temporal/spatial gain>=.01 AND both gains>0 AND >=7/10 spatial folds gain>0 AND both conditional interval lower bounds>0 AND no unsupported primary held-out rows AND benchmark/length/counted-route temporal gains>0. Baseline-form diagnostics apply same numerical screen but are never new-feature discoveries.",
  "uncertain": "Estimable features with some positive categorical temporal/spatial gain OR primary q<=.05, but missing any retention criterion. Null/reversed performance retained.",
  "unsupported": "Tested features with neither positive categorical temporal nor spatial gain and q>.05. Means unsupported by this screen, not no underlying association.",
  "untested": "Six diagnostic-only candidates remain unbanded; 18 deferred remain deferred. Both carry uncertain/not-tested status, excluded from testing families.",
  "promotion": "No Phase6 joint index, coefficient averaging, pricing factor, causal interpretation, residence-risk conclusion or insurer validation. Calibration, correlated features, future source vintages and provisional exposure remain open gates."
}