20mph Evidence Lab — Phase 3 / v3 — 4 November 2026 Release outcome: exploratory controlled comparisons completed. The causal gate is NOT met. Main all-road collision association: -9.75%; nominal 95% calendar-HAC interval -14.76% to -4.44%. Pooled England comparison: -3.07%, with interval including zero. Neither is a proven policy effect. Pre-policy relative trends, poor pre-policy holdout fit, historical placebo breaks and contemporaneous English20mph changes prevent secure attribution. These data do not establish insurer claims, premiums or cost-benefit impacts. OFFLINE REPRODUCTION Download reproduce.py, monthly-panel.csv, annual-controls.json and analysis-specification.json from this directory into one folder. Use Python3.12 with NumPy and SciPy installed, then: python reproduce.py --input-dir . --output-dir reproduced The script verifies pinned input SHA256 values and rebuilds results.json plus all model/chart CSVs. No network is required. Download checksums.json to verify the rest of the release. Numerical floating-point tolerances may vary across library/platform versions; point estimates should agree to at least six decimals. Build environment versions are in environment.json. DATA 132 months, January2015–December2025; Wales plus nine English regions; all-road and recorded20/30mph category counts. Each collision contributes once to collision count and its declared number_of_casualties to total casualties. The original2015–2025 lines from the full DfT dataset are archived in the versioned site source; source and retained-line hashes are in audit.json. The acquisition and aggregation scripts are also in that source. The public monthly panel is the sufficient model input, not a replacement for original record-level provenance. National geography uses ONS district location, not reporting police force. English regions use current highway authority, fallback to source highway authority when current is missing; district country wins in conflicts. The legacy Sheffield code is mapped to Yorkshire and The Humber. Missing coordinates do not exclude records. The public geography-lookup.json documents the map. 158/220 RAS0402 annual checks match exactly; all40 national checks match.62 regional checks differ by at most6 collisions or7 casualties, under0.1% per cell. The mapping/vintage discrepancy remains unresolved. A separate sensitivity scales English region-year totals to official controls while preserving monthly profiles; it changes the main collision estimate by less than0.01 percentage points. The raw panel is never silently corrected. All240 phase1 Welsh monthly collision/casualty checks match. DfT2025 Wales totals are3,004 collisions/4,078 casualties. Later StatsWales totals are3,005/4,079. The recorded20/30 category discrepancy remains18 collisions/23 casualties. Phase2 preserves StatsWales and provisional2026 independently. MODEL Pre:2015–2019 plus January2022–August2023 (80 months). Post:October2023–December2025 (27). September2023 excluded;2020–2021 excluded in main model. Calendar time keeps the two-year gap. Counts measure total reported burden, not exposure-adjusted risk. All series are divided by their own2015 monthly mean. English weights are nonnegative and sum to1; minimize pre-policy collision-index mean squared error plus0.01 times sum((weight-1/K)^2). No post-policy observations choose weights. Casualties retain collision-derived weights. A separate2015–2019 fit predicts January2022–August2023 for a historical holdout check; this is not an unseen prospectively registered holdout. OLS: log(Wales index / weighted English index) = intercept +11 month indicators + calendar-year linear trend + post indicator. Main percentage contrast is100*(exp(post coefficient)-1). Modelled absolute contrast sums post fitted counts minus fitted counts with policy-period terms set to zero, with pre-residual smearing correction. This is not observed or causal prevented events. Headline intervals: calendar-lag Newey–West, max lag12, n/(n-k) finite-sample factor, normal1.96 multiplier. They are nominal conditional intervals, not exact small-sample inference. Secondary intervals: stationary residual bootstrap with expected block length12, geometric restarts, run-centred residuals within each uninterrupted included calendar run; fixed comparison weights. Seed20261104;1,999 main and999 sensitivity draws. Short runs provide limited information; neither interval covers selection, policy contamination or confounding uncertainty. Post-slope sensitivity averages its fitted level+slope contrast. Its interval is bootstrap, not the immediate level coefficient HAC interval; no level-coefficient MDE is reported for this different estimand. Other conditional MDE values use(1.96+0.841621)*bootstrap SD of log level coefficient and are not causal design power. Fake interventions2017-01,2017-10 and2018-10 have12-month follow-up and pre-only refitted weights. They are falsification diagnostics, not independent randomized experiments or a calibrated permutation test. The actual first12-month result is shown on the same follow-up horizon. INTERPRETATION AND VERSIONING No dated treatment-road inventory or verified policy-clean comparison is available. Weights balance outcome histories, not observed covariates. Unmeasured differential changes, police reporting, traffic, enforcement, weather and trend breaks remain. All-severity outcomes avoid direct severity recoding, but not general reporting bias. Recorded20/30mph is a proxy and cannot establish a fixed treated network. No subgroup, injury-severity, mileage-rate or economic claim is made here. The specification is retrospective, written before these comparison estimates but after published Welsh outcomes were known. Later audit/review amendments are explicit in analysis-specification.json: annual-control sensitivity and replacement of a degenerate fixed-length bootstrap with stationary bootstrap plus HAC headline intervals. No favourable specification was selected after seeing results. All planned sensitivities are published. Sources: DfT road safety open data, RAS0402 and linked primary policy sources. source-ledger.json records date precision and download limitations. DfT data under Open Government Licence where stated; policy documents retain publisher rights. Frozen phase1 and phase2 inputs/results remain unchanged. Phase4 subgroup analysis and phase5 economics are not part of this release.