20mph Evidence Lab — phase4/v4 — 4 November2026 Interpretation 21 broad subgroup models pass the declared support rules. These are exploratory all-severity casualty associations, not identified policy effects. Phase3 causal gate remains unmet. Unknown/excluded categories remain visible. Bonferroni intervals cover the21 weighted-model subgroup estimates simultaneously, conditional on the model; the pooled-England sensitivity has a separate21-estimate family using the same critical value. They do not claim joint95% coverage across both models. Holm-adjusted family difference tests are exploratory and not causal. Core offline reproduction Download reproduce.py, monthly-subgroups.csv, comparison-weights.json, analysis-specification.json and input-pins.json into one directory. With Python3, NumPy and SciPy installed: python reproduce.py --input-dir . --output-dir reproduced This verifies input SHA256 and reproduces results.json, estimates.csv, annual-counts.csv, heterogeneity.csv and missingness.csv. results.json contains exact quarterly chart data. Output CSV values carry unrounded precision. Monthly casualty counts exhaust every family; collision counts exist only for collision-level families, and are blank for age/road-user groups to prevent double counting. Source-to-panel reconstruction The versioned site source retains scripts/fetch-20mph-phase4.py and scripts/aggregate-20mph-phase4.py, both acquisition manifests and exact original retained2015–2025 collision/casualty lines in ordered gzip parts. Concatenate each manifest's archiveParts in order, validate every part and combined hash, then decompress; other source years were hashed during retrieval but not retained. Run aggregate script from repository. It validates every casualty-to-collision count, all1320 phase3 monthly totals and10560 exhaustive family totals. DfT originals use OGL. Pin fresh data as a new version; do not silently replace this vintage. Deprivation Comparable England/Wales2019 income release measures2015–16 conditions. Fixed thresholds are below10%,10–below20%,20% or more. The independent ONS exact-fit crosswalk retains unchanged and uniquely split2011parents, and all their children, for every year. Merged/complex parents and children are excluded throughout. Split children inherit whole-parent context. This is collision-neighbourhood deprivation, not casualty household income or exposure. Full source ODS, official crosswalk, stable lookup, extraction scripts and annual coverage audit are retained. No arbitrary post-only missingness correction. Road users and recording Pedestrian0; cyclist1; motorcycles2/3/4/5/23/97; car/taxi8/9; known others retained; unknown−1/99. E-scooter flag goes to other. Official codebook duplicates23; STATS20 and conversion tab establish23electricmotorcycle and33personalpoweredtransporter. No granular serious/slight interpretation without validated adjustments. Fatalities are separate annual counts. Unanswered scope is explicit Adjacent-road diversion cannot be identified without complete dated treated geometry and a fixed buffer/exposure cohort. Speed categories exhaust the network but are not adjacent roads. Wrexham: five sites, July2025 vs two weeksSeptember2025, speed observations only. Edinburgh: historical GIS inspected; paper and workbook arithmetic reconstructed; legal history, matched microdata and causal intervals not replicated. edinburgh-check.json retains the unresolved Table1 rounding/input discrepancy. Raw GIS is linked to the authors rather than publicly rehosted without verified rights. Phase3 remains at /data-lab/20mph/phase3. Phase5 economics is not implemented or implied by this release.