{
  "version": "phase6-v1",
  "frozenBeforePhase6Fits": true,
  "scope": "Retrospective authority all-road screening; not a test of LSOA D1–D10",
  "baseline": "Unchanged phase5-v1 regional_density Poisson specification, cohort, folds, ridge and forward-year rules. M0 replays frozen coefficients; augmented models refit the same baseline terms plus frost using development outcomes only.",
  "cohort": "The same 130 complete authorities and exposure as phase 5; eligibility is retrospectively known over 2021–2025",
  "years": {"development": [2021, 2022, 2023], "assessment": [2024], "without2021": [2022, 2023], "finalYearUsed": false},
  "candidates": [
    {"id": "M0", "label": "No frost", "form": "none"},
    {"id": "M1", "label": "Fixed authority bands A1–A10", "form": "categorical"},
    {"id": "M2-linear", "label": "Continuous frost · linear", "form": "linear"},
    {"id": "M2-curve", "label": "Continuous frost · restrained curve", "form": "spline"},
    {"id": "M3-rural", "label": "Frost × resident rural share", "form": "rural"},
    {"id": "M3-trunk", "label": "Frost × trunk mileage share", "form": "trunk"}
  ],
  "features": {
    "frost": "Frozen phase4-v1 authority road-length-weighted air-frost normal, 1991–2020, and fixed authority bands; no new quantiles for categorisation",
    "categorical": "A1 reference and A2–A10 indicators; fixed ridge coefficient 1 after training-only standardisation, matching other continuous design terms",
    "curve": "Restricted cubic spline with 3 knots at training-authority frost quantiles 0.1, 0.5 and 0.9 (2 degrees of freedom including linear); linear beyond boundary knots; knots recomputed inside each training fold",
    "interactions": "Linear frost main effect plus one interaction. Rural share is 1 minus resident urban share. Trunk-mile share is a road-mix proxy, not detailed road class. Centre frost and modifiers in training only.",
    "noWeatherOrSurfacePredictor": true,
    "M4": "Deferred to phase 7; no adjacent merging in this release"
  },
  "selection": {
    "folds": "Exact phase5-v1 folds, keeping whole authorities and repeated years together",
    "score": "50% forward-year mean Poisson deviance plus 50% geographic mean Poisson deviance, identical rows for every candidate",
    "temporal": "2021 predicts 2022; 2021–2022 predict 2023. Dynamic covariates and calendar effect carried from latest training year.",
    "spatial": "Five within-region geographical folds, same-year non-outcome covariates, distinct from future-year prediction",
    "parsimonyTolerancePercent": 0.1,
    "tieBreak": "Fewest mean parameters within 0.1% of best development score, then smallest score, then candidate id. Includes M0.",
    "heldOutResultsCannotChangeSelection": true,
    "baselineSpecificationWasPreviouslySelectedOnDevelopment": "Conditional extension comparison, not fully nested joint baseline/frost selection"
  },
  "assessment": {
    "primary": "2024 predicted-mean Poisson deviance on identical rows and realised motor-mile offsets; no held-out intercept or year recalibration",
    "referenceThresholds": "Reuse v1: positive 2024 gain; paired interval lower bound above zero; improvement in at least 4 of 5 spatial folds; overall O/E 0.95–1.05 and adequately supported regional O/E 0.90–1.10. The 0.5% final-year threshold belongs to phase 8 and is not claimed tested here.",
    "uncertainty": "1000 paired region-stratified spatial-block resamples of regional phase5 fold groups. Keep all observations within each sampled block; fixed fitted models, no refitting. Conditional on model selection, exposure and block scheme; not uncertainty over future weather years.",
    "effectUncertainty": "Pointwise t-based sandwich intervals with region × geographic-fold score clusters and finite-cluster correction; conditional on selected specification and regularisation. Show only regional observed frost range. Not Poisson standard errors or individual risk factors.",
    "effectReferenceDays": 35,
    "without2021": "Refit each same specification using 2022–2023 only; no reselection"
  },
  "weatherDiagnostics": {
    "annualAnomaly": "Blocked: matched annual air-frost observations have not been acquired. Registered CEDA access is required for the intended release. Do not replace air frost with monthly mean minima, ground frost, collision-surface counts or fixed normals.",
    "futureAnomalyDesign": "Separate location and year effects with annual frost minus the fixed normal; contemporaneous association unless deployment-time weather forecasts are supplied",
    "secondaryOutcome": "Same eligible motor-involved collision IDs, road_surface_conditions=4; dry/wet/snow/other valid codes separate from -1/9/blank missingness. Reconcile all eligible collision counts to phase5 panel before modelling.",
    "secondaryModels": "M0 and predeclared M2-linear on frost/ice counts, trained 2021–2023, assessed 2024. No candidate selection and no promotion to primary outcome.",
    "winter": "December, January and February collision burden per annual motor-mile; never winter frequency or frost-day risk. No daily exposure estimate.",
    "missingness": "Annual and regional missing-surface percentages and counts; missing codes never treated as known non-ice. More than 5% missing flags secondary interpretation. No imputation."
  },
  "interpretation": "Exploratory incremental screen while phase5 scientific gates remain open. 2024 outcomes and 2025 descriptive frost associations were already inspected in earlier phases; this is not untouched or prospective validation. No causal, insurance-claims, home-postcode or all-road LSOA conclusion."
}