{
  "version": "phase7-v1",
  "date": "2026-10-11",
  "frozenBeforePhase7Fits": true,
  "scope": "Phase 7 authority-band simplification, conditional on the frozen phase 5 baseline. A1–A10 are authority bands, not national LSOA D1–D10.",
  "developmentYears": [2021, 2022, 2023],
  "assessmentYear": 2024,
  "finalYearUsed": false,
  "inheritedThresholds": {
    "rateRatioBounds": [0.95, 1.05],
    "maximumMergeDevianceLossPercent": 0.1,
    "overallCalibrationBounds": [0.95, 1.05],
    "regionalCalibrationBounds": [0.9, 1.1]
  },
  "equivalence": {
    "method": "Every pair of original bands inside a proposed contiguous group must have its entire two-sided 95% cluster-sandwich adjusted-rate-ratio interval inside 0.95–1.05. A non-significant difference is insufficient. Requiring every within-group pair prevents chains of individually small differences from creating a wide merged group.",
    "clusters": "Inherited region × geographical-fold blocks; repeated authority-years stay together. Finite-cluster correction and t(G-1) quantile. Conditional on the original ten-band fit and its fixed ridge, not simultaneous post-selection confidence or proof of equivalence in a new population.",
    "minimumAuthoritiesPerOriginalBand": 5,
    "minimumBlocksPerOriginalBand": 3,
    "minimumTrainingCollisionsPerOriginalBand": 30,
    "missingBand": "A missing or unsupported band cannot be declared equivalent.",
    "candidateSet": "All 512 contiguous partitions of the ten original bands, including ten unchanged bands and one group. Prune only by the declared support/equivalence gates; never cut at a new frost value.",
    "innerTrainingProtection": "A proposed merge must satisfy equivalence and support in the complete selection-training set and in every inner training fold. Inner validation outcomes never enter that fold's effect estimates. Original cutoffs were frozen without outcomes."
  },
  "selection": {
    "outer": "Repeat the complete representation-selection procedure inside each of the existing two forward-year and five geographical development splits. No outer held-out outcomes enter inner selection. This nests band/representation selection, conditional on a baseline specification previously selected on the same development era.",
    "inner": "Within each outer training set, use its available forward-year splits and its available original geographical folds. Keep whole authorities together in spatial splits. With only one training year, use geographical validation alone; do not borrow a later year.",
    "score": "Equal weights for temporal and spatial mean Poisson deviance when both exist; spatial mean deviance when forward-year validation is unavailable.",
    "comparators": ["M0", "M1", "M2-linear", "M2-curve"],
    "candidatePenalty": "Unchanged phase 6 Poisson baseline and training-only centring/scaling. One indicator per non-reference merged band, fixed ridge 1. No tuning against assessment outcomes.",
    "decision": "Eligible simplifications must be within 0.1% of both the best inner score and unchanged M1, and must not worsen calibration relative to M1. Choose the fewest full-training mean parameters, then score, then deterministic ID; otherwise retain M1. M0 can collapse all bands without asserting equivalence when it meets the same predictive rule.",
    "calibration": "For each available validation type, overall and regions with at least 10 authorities: if M1 meets the inherited O/E bounds, a simpler representation must also meet them; if M1 misses them, the simpler representation must be at least as close to 1 in absolute log O/E. No held-out recalibration. Passing this relative screen does not close the absolute scientific calibration gate.",
    "tail": "A9 and A10 obey exactly the same equivalence, support and performance gates. There is no forced tail merge and no forced monotonicity.",
    "freeze": "Select the final representation using only 2021–2023 inner validation; write coefficients, exact mapping, evidence for all pairs, and row-level split membership before loading 2024 assessment observations. Save the chosen mapping for phase 8."
  },
  "assessment": {
    "primary": "Unchanged 2024 rows and realised authority mileage offsets, with dynamic covariates and calendar effects carried from 2023. Compare frozen selected representation, M0, M1 and both continuous models.",
    "without2021": "Refit the same frozen mappings/specifications on 2022–2023; no reselection.",
    "uncertainty": "1000 paired, within-region geographical-block resamples of fixed assessment predictions, as in phase 6. Conditional on model choice, exposures and block definition; not allocation, selection or future-weather uncertainty.",
    "2025": "No 2025 fitting, scoring, selection or persistence claim. Descriptive 2025 results were viewed in earlier phases; phase 8 will be retrospective, with prospective confirmation requiring a new untouched year."
  },
  "lsoaMileagePilot": {
    "regions": ["London", "South East", "South West"],
    "release": "mileage-v1",
    "bandMappingChanged": false,
    "reason": "All-road LSOA exposure is still conditional, with no genuinely new 2024 count sites, frost-stratum bias failures and no independent validation of within-authority allocation. Authority band results cannot establish LSOA merges. Preserve the existing modelled-mileage sensitivities and strict major-road subset separately; do not certify any D-band simplification from this authority phase."
  },
  "interpretation": "Complete the supported authority phase-7 analysis even if no merge is justified. This does not close the LSOA, annual-weather, baseline-calibration or phase-8 persistence gates. Prior 2024 and 2025 descriptive inspection is disclosed; this is retrospective evidence, not prospective validation."
}