20mph Evidence Lab — Phase 1, version 1 Frozen: 3 November 2026 https://johanstrydom.com/data-lab/20mph What is complete Source register, refreshed 2025 STATS19 ingestion, validated 2021–2025 joins, Welsh annual reconciliation, TfW file checks, five dated policy/history evidence records, a road-section schema, preliminary comparison screening, illustrative power sensitivities, and a versioned study protocol. No policy effect is estimated. Decision Conditional go for a Phase 2 area-level descriptive pilot after historical import and stable-geography checks. Hold road-level causal work: verified section/date pairs, road exposure and uncontaminated controls are absent. Coverage percentage is unknown. Zero rows in road-sections.csv means no verified study sections. Reproduce without access to the site repository Download reproduce-20mph.py and run with Python 3.10 or later: python reproduce-20mph.py NEW_DIRECTORY It creates a new directory, downloads approximately 60 MB of archived inputs and publisher files, checks SHA-256 hashes, runs both audits, and compares their exact output hashes. Network access is needed for the initial downloads. No Python packages, API key or paid data is required. Changed publisher bytes fail closed. The reproduction script trusts this published v1 manifest, not an external signature. Fresh 2025 files are reconstructed from the existing five-year archive and compact line-order/delta files. Every reconstructed original SHA-256 must match. This removes redundant public copies without changing any input bytes or calculations. Reproduce in the site repository, entirely offline python scripts/audit-20mph-monitoring.py python scripts/build-20mph-phase1.py python tests/20mph-phase1.test.py Inputs: original STATS19 archives, byte-exact refreshed 2025 CSV reconstruction files, frozen geography extract and TfW ODS snapshots. Original DfT data are OGL v3.0. TfW dataset-specific reuse licence was not verified; the public source register links publisher files. The private repository archive is retained for reproducibility, not mirrored publicly. Reading the files - audit.json: checks, reconciliation, coverage and decision; no treatment estimate. - official-controls.json: manually transcribed Welsh publication controls and date. - source-register.json: per-source origin, vintage, hashes, status and limitations. - protocol.json: primary outcomes, periods, comparisons, sensitivities and gates. - wales-monthly-audit.csv: country from ONS district, all roads and recorded20/30. This is an import/reconciliation extract. The categories are not a fixed network. - district-month-audit.csv.gz: diagnostic grid from observed district identifiers. Boundary changes can create structural zeros. Not approved as a causal panel. - comparison-area-screen.csv: pre-policy trajectory similarity only; no approved controls. 2016–2019 and2022, not the full intended2015–2019 baseline. - power-sensitivity.csv: approximate count-based detectable reductions, conditional on hypothetical control volumes and variance inflation. This is not validated power for one national intervention. See audit.json for equation and assumptions. - road-history-evidence.json: exact commencement, proposal dates, effective-by bounds and observed limits are distinct. Dates never inferred from map stages. - road-sections.csv: header-only register until exact legal dates and geometry pass. - monitoring-audit.json: speed table discrepancy and120 journey consistency rows. - manifest.json: study-file checksums; source-register hashes identify original bytes. Material findings to retain Welsh2025 totals3004 collisions and4078 casualties match the June2026 publication. Recorded20/30 collision count1488 differs from the published1469 by19. Vintage or classification is a hypothesis, not a resolved explanation. The archived September and refreshed November2025 collision extracts are compared record by record. TfW Summer2025 <=24mph KPI51.5% differs from Table4 bins54.0%; Table4 is quarantined. Positive whole-day ODS offsets are removed only for known sub-day journey durations; raw ODS remain intact. All three differences in each of120 rows reconcile to0seconds. Recorded serious injuries are reconciliation controls, not a comparable trend. Next acquisition priorities (no messages or data requests have been sent) 1. Full2015–2020 STATS19 microdata and vintage-aware annual reconciliation. 2. Stable district/authority geography for the full period; explicitly bridge changes. 3. Welsh authority section geometry with legal commencement/end dates, pilot, exception, traffic-calming, part-time and reversal classifications. 4. Resolve the19-collision source difference and TfW distribution-table conflict. 5. Verify English candidate-area policies before freezing a control pool. 6. Match annual road exposure with counted/estimated flags and quantify missingness. 7. Confirm TfW reuse terms before mirroring files or publishing extensive extracts. Amendments must create v2 with a reason; do not silently overwrite this frozen release.