{
  "version": "phase8-v1",
  "date": "2026-10-12",
  "frozenBeforePhase8OutcomeAssessment": true,
  "primary": "Replay phase7-v1 selected authority model M1 and M0, without refitting or recalibration, on the same 130 authorities in 2025. Realised official aggregate motor miles are the offset; dynamic covariates and calendar effects remain at 2023. Do not substitute a better sensitivity after inspecting the final year.",
  "priorInspection": "2024 and descriptive 2025 outcomes were examined in previous phases. This is a frozen retrospective assessment, not an untouched prospective test. Prospective confirmation needs a new complete year.",
  "thresholds": {
    "minimum2025DevianceImprovementPercent": 0.5,
    "positive2024Gain": true,
    "paired95PercentLowerBoundAboveZero": true,
    "minimumPositiveSpatialFoldsOfFive": 4,
    "overallObservedExpected": [0.95, 1.05],
    "regionalObservedExpected": [0.9, 1.1],
    "minimumSupportedRegionalAuthorities": 10,
    "tailSameDirectionYearsOfFive": 4
  },
  "authoritySensitivitySet": {
    "without2021": "Replay phase7-v1 2022–2023 fits, retaining the selected representation, on 2024 and 2025.",
    "rolling": "Fixed M0/M1 specification: 2021 predicts 2022, 2021–2022 predicts 2023, 2021–2023 predicts 2024, and 2021–2024 predicts 2025. The last is a separately labelled rolling refit, not the primary phase 7 coefficients. All fits are frozen before loading 2025 outcomes.",
    "geography": "Refit fixed M0/M1 on 2021–2023 after excluding each of the original five within-region authority folds; predict all years on excluded authorities. Repeat using four outcome-free recursive centroid folds within each region. Shared authority-years stay together. Regional fold boundaries can still share climate cells/roads, so spatial independence is approximate. Model-specification selection was not repeated: report this conditional fixed-specification test separately from phase 7 nested selection.",
    "outOfRegion": "The original regional fixed-effect model cannot predict an unseen region. Test transportability separately using pooled regional_density controls with region intercepts and London/South East density deviations removed, retaining density squared, road/traffic controls and calendar terms. Fit on eight regions in 2021–2023, predict the ninth in 2025. No outcome-derived intercept for the held-out region. Report frost-range and band-support extrapolation. This is a different sensitivity specification, not primary-model validation.",
    "areaWeighting": "Use the existing authority area-weighted air-frost normal; create outcome-free equal-authority nearest-rank cutoffs over the phase 4 reference frame, requiring at least 98% aggregate valid area coverage and a finite area normal. Publish coverage exclusions and compare both models on the identical restricted cohort. Ties remain together. Fit the unchanged categorical form on 2021–2023. Publish separate area-band cutoffs; do not alter road bands or LSOA D1–D10.",
    "uncertainty": "1000 paired, within-region original geographic-block bootstrap draws; comparisons share the same draws within year/scope. Repeat primary uncertainty with whole-authority blocks. Fixed predictions; no refitting. These intervals omit model-selection, exposure-error and future-weather uncertainty.",
    "groundFrost": "Unavailable in the pinned input releases; no air-versus-ground comparison is claimed.",
    "annualWeather": "Realised annual air-frost observations remain unacquired. No weather-anomaly persistence or frost-day exposure is inferred."
  },
  "tailAnalysis": {
    "authority": "Fixed original A1–A10; report all five years of counts, motor miles, raw rates, baseline O/E and final-model O/E. Use original five-fold spatial out-of-fold predictions for tail diagnostics; authority outcomes are excluded from their prediction fit, but specification selection used the development era. In 2021–2023, calendar effects come from peer areas in that year; 2024–2025 carry 2023 covariates/effects. Keep this distinct from primary full-fit out-of-time calibration.",
    "contrasts": ["A10 / A9", "A9 / A8", "A10 / A8–A9"],
    "adjustedRatio": "(observed / M0 expected) in higher group divided by (observed / M0 expected) in comparison group; also publish the M1 calibration ratio separately. This residual contrast is not a fitted frost coefficient or causal effect.",
    "support": "At least five authorities, three original region-by-fold blocks and 30 collisions in EACH side/year. Otherwise flag limited support and withhold confidence intervals. Bootstrap intervals require at least 800 finite draws of 1000.",
    "withinRegionAndRoadContext": "Report regional contrasts and regional strata using frozen resident urban share >=0.8 versus <0.8, and above/below the cohort median 2023 trunk-mile share. Resident urban share and trunk management are limited road-environment proxies. Sparse strata are inconclusive.",
    "persistence": "Count years with adjusted ratio above/below one and compare with the original four-of-five direction rule. Require all five years supported for an unqualified persistence pass. These five annual contrasts are dependent; do not treat them as five independent weather experiments. No requirement that every annual interval exclude one."
  },
  "lsoaPilot": {
    "scope": "Separate conditional LSOA D-band evidence for London, South East and South West, 2021–2024 only. No 2025 exposure exists in mileage-v1 and no validation gate is relaxed.",
    "allRoads": "Reuse the frozen mileage-pilot preferred, benchmark and +/-10%, +/-20% exposure assumptions and the existing 2024 comparisons, with provenance. Add annual D-band count/exposure summaries; do not label raw trends as incremental predictive persistence.",
    "majorRoads": "Compare the existing balanced strict routed subset with a smaller common counted network. Start from routed sections, reject every section sharing an OS edge with another section in a year, require Counted in all four years and identical routed OS-edge sets across all four years. Require a unique candidate count point for each year/LSOA/road-number combination, and retain count-point/LSOA pieces present in all four years. This reconstructs count-point attribution only where unambiguous from the archived allocations; no new nearest-road matching is invented.",
    "countedNumerator": "Use only previously matched-once collisions. Attribute a count point only when the archived accepted section allocation has exactly one candidate for that collision's year, LSOA and road number. Exclude other collisions and their corresponding ambiguous exposure pieces. Preserve zero-collision positive-exposure pieces. Restrict to the original balanced eligible LSOA cohort.",
    "countedComparison": "Refit the pilot's fixed Poisson M0 (region, log1p density, resident urban fraction) and M1 (adds original national D bands), same ridge and offset, on 2021–2023. Assess 2024 only, including a without-2021 sensitivity. Report full routed and counted network scopes separately because they contain different roads. This does not validate 2025 persistence.",
    "limits": "The climate score is the original all-mapped-road LSOA normal, not recomputed specifically on counted major sections. Road matching, network scope and within-LSOA exposure remain conditional. Counted AADF is a traffic estimate based on counts, not a census of annual travel. No LSOA rating factor is promoted."
  },
  "completion": "Complete supported authority phase 8 analysis and publish pass/fail results. Record missing ground/annual frost, independently validated LSOA exposure and 2025 pilot mileage as unresolved workstreams. Phase 9 remains the consolidated publication and final evidence handoff; publishing this phase's results does not claim every original study scope is complete."
}