England and Wales national feature expansion ew-v3 · E4–E7 · 12 October 2026

The full-feature-decile-lookup.csv.gz provides every master LSOA and all 62 candidates in one table. The smaller incremental lookup is also supplied.

13 E4–E7 candidates on all 35,672 rows. Join v1 and v2 by lsoa; replace their 12 E4–E7 placeholder columns with v3 and add heat_summer_days_gt25. 62 candidates total.

One England-and-Wales master, equal LSOA weights. Source coverage alone governs eligibility. Never rerank regional/insurer samples. Partial environmental references are explicitly provisional; crime and speed national bands withheld.

2025 air/insolvency and 2026 roads/crime editions are retrospective context for prior collision years, not historical as-of forecasts. No prospective 2022 claim.

OGL for official sources; OSM-derived speed database under ODbL 1.0 with © OpenStreetMap contributors. A combined collection does not relabel source licences.

E4: supporting >25°C observed 1991–2020 normal and assessed 2025 NO2/PM2.5; 2022–24 air sensitivity separate. Exact inclusive heat thresholds are source-access blocked.
E5: own-territorial-force observed rates only. All national crime bands withheld. Month coverage, unlocated records, duplicate/conflicting IDs and other-force/BTP diagnostics retained. Census 2021 denominators are fixed proxies.
E6: dated OSM speed shares, >=95% classified-length gate, unknown length and lower/upper bounds. Only 1,547 areas qualify, so primary national speed bands are withheld. No default-speed imputation; authority validation pending. DfT count coverage is audited; no national flow deciles claimed.
E7: published 2025 bankruptcy and total insolvency authority context, 2024 adults. Shared rates/ties retained; no invented within-LSOA detail. 2015–25 history and authority sensitivity supplied.

Offline reproduction: unzip reproduce-ew-v3-e4-e7.zip, run python scripts/build-ew-e4-e7.py --out-dir regenerated. Compare every output SHA256 in manifest.json. Compact source aggregates reproduce the registry; deep source extraction uses the pinned original downloads and separate acquisition scripts. No accident outcomes or claims are inputs.
Downloads served as exact, hash-checked frozen bytes. No prediction or causal effect established. E8 final freeze and modelling remain open.