Social decile validation v2 — Phases 5–7

EXPLORATORY. No decile claim passes every registered gate. No causal, driver-level or pricing claim. All 2023–2025 outcomes had already been inspected. The original v1 artifacts remain unchanged.

REPRODUCE THE MODEL VALIDATION
Unzip reproduce-v2.zip into a new directory. Python dependencies: numpy and scipy at the versions recorded in frozen-design.json. Set OPENBLAS_NUM_THREADS=1.
1. python scripts/social-deciles-validation-amendment-v2.py correct-fit
2. python scripts/social-deciles-validation-v2.py bootstrap --workers 4
3. python scripts/social-deciles-validation-amendment-v2.py report

The bundle contains all SHA-checked inputs needed by the extension, including the original v1 rows, fits and model code. It starts at the recorded amendment state, with the first attempt's geographic artifacts. The correction converts an absent training-vintage cutpoint into an unavailable stress fit; it neither borrows assessment cutpoints nor changes an estimator. All other original designs remain frozen. The uncorrected original fit attempt and its error are documented in numerical-amendment.json.

999 seeded regional bootstrap replicates refit all six models across the three temporal folds. Checkpoints are resumable by replicate ID. Worker count changes scheduling only. Report creates a gzip of complete replicate results. For the web release the same JSON bytes are compressed as bootstrap-summary.json.xz to fit the hosting size limit. Read with Python lzma.open or an XZ decompressor. Every fit diagnostic and failed/undersupported case is retained. Point predictions and bootstrap numerical values are deterministic; run timestamps differ.

Intervals condition on realised traffic and contain only nine independent spatial blocks. Bootstrap sparse/rank-deficient designs remain archived with the fixed numerical ridge. Point-fit support failures are separate acceptance-gate failures. Geographic transfer scores are point estimates; the registered refitted uncertainty concerns paired temporal comparisons.

CHARTS AND FINDINGS
predicted-rates.csv/JSON aggregate the same held-out counts and miles. Model-adjusted means conditional held-out prediction at the observed covariates, not a standardised causal effect. Display deciles remain fixed when filtering. findings.json is byte-identical to the website rendering copy and drives Data Lab and the independent Lens follow-up card. Manifest hashes pin inputs, outcomes, maps and downloads. Unknown selections fail closed.

PHASE 6
The LSOA exposure pipeline and its input contract are in ../lsoa-exposure-v1. Real exposure gates fail: no complete historical all-road network allocation, independently validated minor-road mileage, matched stratum reconciliation or exposure uncertainty set. Zero eligible LSOA frequency rows; profile motorMiles and collisionRatePer100k remain null. No authority model is transferred to LSOAs.

MAP AND UI LIMITATIONS
The authority social-context map uses exact-code 2025 ONS polygons, simplified for display only. The LSOA coordinate retrieval was blocked by a network-proxy 403; no partial neighbourhood map is published. LSOA profile lookup remains available. Browser interaction/mobile visual QA was unavailable; SSR selection tests, responsive styles, accessible labels/tables, type checks and the production build are checked separately.

RELEASE / ROLLBACK
v2 is an independent study route selection. version=v1 preserves the prior decile release. A site rollback must restore the saved site version as a whole, including finding-register copy, manifests and UI; never mix result versions. A future candidate must use a new protocol version. Phase 8 monitoring extension is not part of this release; the new artifacts are not silently added to the old scheduler. No human review is claimed.