{
  "version": "phase5-v1",
  "date": "2026-10-11",
  "purpose": "Freeze a no-frost authority baseline using development data, then assess 2024 calibration; review London first, South East second, other English regions next.",
  "scope": "Matched all-road authority motor-involved injury collisions and official aggregate motor-vehicle miles. No LSOA or matched-major-road frequency model.",
  "outcome": "The unchanged frost v1 definiteMotorCollisions numerator. No parking model outcomes or fitted results are used.",
  "cohort": "Start with the phase 4 balanced 131 authorities. Require complete primary controls in 2021–2023 and exact mileage-context joins; retain the same eligible authorities for all candidate comparisons and 2024 assessment. No imputation. 2025 does not enter fitting, selection or baseline assessment.",
  "features": {
    "primary": ["urbanShare", "logDensity", "trunkShare", "vanShare", "hgvShare", "logFlow"],
    "static": "ONS 2021 resident urban share and resident population density; resident context is not urban-road mileage.",
    "dynamic": "Trunk-road management and van/HGV motor-mile shares, plus log(annual motor miles / official road-km). No distribution among collision records is used.",
    "forwardPrediction": "Carry each authority's last training-year dynamic controls forward. Keep realised held-out motor miles only as the count exposure offset. Do not use held-out traffic mix, held-out year coefficients or held-out collisions to forecast its baseline rate.",
    "missing": ["Detailed road-class and speed mix", "Elevation", "Coastal distance"],
    "sensitivityOnly": "April 2026 mapped junction density and social features (income deprivation and no qualifications) are evaluated inside development data only. The network snapshot follows outcomes. A parking score is an unvalidated derived proxy and is deferred rather than adopted as an observed control."
  },
  "candidateMeanStructures": {
    "regional": "Region intercepts and training-year categorical effects only.",
    "road_context": "Regional model plus urbanShare, logDensity, trunkShare, vanShare, hgvShare and logFlow.",
    "regional_density": "Road-context model plus squared centred logDensity and London/South East logDensity slope deviations; ridge regularisation limits flexibility."
  },
  "families": ["Poisson", "NB2, Var(Y)=mu+alpha*mu^2; alpha estimated on training data"],
  "fit": "Log link, log(motorMiles/100000) offset. Fixed ridge penalty 0.5*sum(beta^2) on standardised continuous terms; 1e-6 stabiliser on non-intercept region/year terms. Training-only centring, scaling and constant-column removal. No location fixed effects and no frost features, bands, interactions or outcomes by frost level enter fitting or selection.",
  "yearForecast": "An unseen future year uses the last fitted training-year effect, never a coefficient estimated from assessment-year outcomes. Historical source versions are retrospective, so this is temporal validation rather than a strict historical availability backtest.",
  "developmentSelection": {
    "temporal": "Fit 2021 and predict 2022; fit 2021–2022 and predict 2023. Freeze dynamic covariates at each training cutoff.",
    "spatial": "Five deterministic geographic folds within each region, using recursive centroid splits along the widest scaled coordinate and retaining all years of each authority. Held-out geography uses contemporaneous non-outcome covariates: spatial transfer is reported separately from temporal forecasting.",
    "score": "0.5*temporal Poisson deviance per held-out authority-year + 0.5*spatial Poisson deviance per held-out authority-year. All candidates use identical rows.",
    "parsimony": "Choose the fewest mean parameters within 1% of the minimum development score; then lower score, then Poisson in an exact tie. Numerical failures disqualify a candidate, without changing the candidate set.",
    "refit": "Refit the selected candidate on 2021–2023, write frozen coefficients and scaling before evaluating 2024 outcome values; source files are loaded for integrity and context, but test outcomes are not used in selection. Do not refit after calibration is assessed."
  },
  "uncertainty": "Regional/overall observed-to-expected ratios have 95% percentile intervals from 2000 resamples of whole authorities. These describe between-authority variation conditional on fitted predictions, omit parameter, mileage and cross-authority spatial uncertainty, and are not causal intervals.",
  "diagnostics": {
    "calibration": "Overall 2024 O/E within 0.95–1.05; regional within 0.90–1.10, using previously frozen frost protocol tolerances. At least 10 authorities for an unqualified regional calibration flag. Report failures without tuning.",
    "dispersion": "Training Pearson residual sum of squares / residual degrees of freedom under the fitted family; also show its Poisson-variance counterpart and NB2 alpha.",
    "spatialResiduals": "Moran's I on authority Pearson residuals, symmetric four-nearest-centroid graph, 999 permutations. England residuals are centred within region and permuted within region; regional tests are centred/permuted within region. Two-sided permutation results are exploratory, with multiple comparisons disclosed.",
    "contrast": "Report fixed authority frost range, populated national A bands, bands with at least five authorities and A9/A10 support. Continuous contrast flag needs at least 10 authorities and at least 10 frost days of range. Categorical support flag also needs at least two A bands with five authorities. These are descriptive design thresholds, not evidence that frost predicts collisions.",
    "without2021": "Refit the same selected specification on 2022–2023, without reselecting it; assess 2024 separately.",
    "gate": "Separate implementation completion, numerical validity, calibration and frost contrast. Do not call the scientific phase gate passed where calibration or support is inadequate. Significant residual spatial structure is an open diagnostic, not proof of a missing frost effect."
  },
  "holdoutDisclosure": "Phase 4 already displayed 2025 frost associations. This phase leaves 2025 out of its modelling operations; it does not restore an untouched-test claim. Future prospective confirmation needs a new untouched year.",
  "status": "Design frozen before phase 5 model fitting; later stages must preserve this release and use a new version for changed choices."
}