England and Wales additional features — F1–F6 — ew-f6-v1
Frozen 13 October 2026

SCOPE AND RESULTS
The entire 35,672-LSOA2021 master is retained: England 33,755; Wales 1,917.
Of eleven targeted additions, 2 receive national bands, 4 have diagnostic values only, and 5 remain unavailable under the frozen target definitions.
The combined registry has 40 banded features, 10 diagnostics and 12 deferred candidates, out of 62. The original 38 banded features and their Phase 5 results are unchanged. None of the additions has entered F7 collision testing. No insurer-claims validation or causal proof is claimed.

USE THE NATIONAL LOOKUP
full-feature-decile-lookup.csv.gz contains every master row and all 62 feature value/band/status triples. Join by lsoa; never average deciles or rerank a local subset. Existing non-target columns are preserved exactly from the prior lookup. The 621 prior public files remain byte-identical.
D1 means low source value, D10 high. Where declared, Z0 is observed zero and P1–P10 rank positive values. Identical values stay together; empty and unequal bands are legitimate. Blank is missing or withheld, never zero or a middle band. Claimant change can be negative. Numeric diagnostics do not receive invented bands.
Source observations, gridded estimates, terrain proxies and unavailable values have explicit statuses. Current snapshots and October 2026 band definitions are retrospective; they are not certified available at a 2022 forecast origin.

DECISIONS AND SOURCE LIMITATIONS
phase-decisions.json and quality-audit.json record the coverage, measurement and missing-rank gates for every target. coverage.csv also reports country, region, rurality and authority counts. Missingness is not assumed random. A matched-cohort and geography sensitivity remains required in F7.
Heavy-rain coastal grid gaps retain 307 unavailable LSOAs; eligible LSOAs require at least 99% covered road length. Rank bounds put all unknown values below and above observed values.
School source records and optional maintained-school calculations are separate from the all-school target: missing locations or incomplete school eligibility are not zero.
Claimant estimates use age16–64, full annual observations and exact historic-geography aggregations only. Split/complex areas are unavailable.
Terrain50 is ground elevation. Numeric along-road terrain proxies and sensitivity checks do not validate road-surface gradient, especially at bridges and tunnels.
Four exact weather definitions require authorised CEDA files. No login bypass, strict-greater-than heat substitution, or total-rainfall/wet-day approximation is used.
NaPTAN access-site density describes physical sites near eligible roads, not service frequency or passenger flow. Current coordinates, active flags and generalised road geometry limit historical interpretation.
See f2-audit.json through f5-audit.json for definitions, official sources, licences, source hashes, data periods and each remaining gate.

REPRODUCE F6 OFFLINE
Download reproduce-ew-f6-v1.zip, extract into a clean directory and run with Python3.10+ (standard library only):
  python scripts/reproduce-ew-f6.py
The runner verifies all bundled input hashes, rebuilds the combined national lookup, national bands, summaries and gate decisions, and checks every numerical output byte-for-byte.
This kit reproduces F6 from frozen source aggregates. It includes the prior master, lookup and summary; the four new long feature tables and source audits; and the exact F1 protocol. It does not repeat every earlier phase or the full national spatial extraction. The prior file-preservation check was executed in the complete project; the compact kit verifies its own required inputs instead.
The acquire/build scripts for F2–F5 are supplied for deeper reconstruction, with source URLs and hashes. They need their documented source files and project geometry helper, and Python geo dependencies (numpy, pandas, geopandas, pyogrio, shapely, rasterio, pyproj, pyarrow; optional netCDF4). The large OS/ONS original geometry and authenticated weather files are not bundled. Public input files and phase audits retain compact extracts and detailed spatial components.

NEXT: F7 COLLISION SCREENING
Use the frozen STATS19 motor-involved injury-collision outcome and matched motor-mile offset log(miles/100000). Compare on identical eligible rows against the Phase4 baseline. Fit2022 and predict2023; leave each of nine English regions and Wales out in turn. 2024 and2025 were previously inspected and remain diagnostic. Preserve the eleven-feature primary testing family, including unestimable candidates, and reserve a complete untouched future period for confirmation. Insurer trials remain separate.

VERIFICATION
release-verification.json records executed tests and limits. Browser interaction checks were unavailable in the managed environment; server rendering, build and live HTTP byte checks are distinct checks.
Source-specific OGL terms remain in force; earlier OSM-derived data retains ODbL attribution. Higher feature values do not imply greater collision risk.
