PARKING AND COLLISION FREQUENCY — PHASE 3 Frozen 10 October 2026 Scope Exploratory authority-level descriptive results, not a causal or adjusted parking effect. England, all public roads, 2021–2025. A fixed 140-authority cohort contains 446,959 police-reported injury collisions. No LSOA mileage is imputed. The original residential-road LSOA test remains incomplete. Reproduction from the Site source checkout 1. Use Python 3, NumPy 2.3.5 and SciPy 1.17.0 (see analysis.json for the actual build versions). No network connection is needed for the builder. 2. Run: python scripts/build-parking-phase3.py 3. Run: python tests/parking-frequency.test.py UI verification: node --test tests/parking-frequency.test.mjs Reproduction from downloads Create a project directory with a scripts/ folder. Save the downloadable build-parking-phase3.py into scripts/. Save input-manifest.json into public/research/parking/v3/. Download each file named in the manifest, retaining its complete relative path. Public files are available from: https://johanstrydom.com/ + the path with its leading public/ removed. The builder checks every input's SHA-256 and length before calculating. Run the builder from the project root. It recreates all v3 analysis JSON and CSV outputs. The manifest itself and frozen authority-context.json are inputs. Python and dependency versions can affect floating-point last digits; analysis.json records the original environment and seed. Key outputs analysis.json: methods, source register, exclusions, annual and pooled results, all 27 model contrasts per measure, no-London sensitivity. authority-year-panel.csv: all 153 x 5 source rows, with eligibility, fixed-cohort membership and reasons. Excluded source mileage can remain present for audit; do not include excluded rows in the analysis. authority-decile-membership.csv: both parking measures x 27 variants, 140 authorities per variant; membership frozen across years. decile-rates.csv: both measures x 27 variants x 6 periods x 2 outcomes x 10 deciles. Source miles are absolute vehicle miles, not millions. scenario-contrasts.csv: D10/D1 ratios and authority bootstrap intervals. data-dictionary.json: units, nulls and grain. Interpretation Reference off-street model, pooled all-injury outcome: D10/D1 is about 0.13. Excluding London and re-ranking changes it to about 0.71. Across 27 model variants the ratio spans about 0.11 to 5.63, reversing direction. These are sensitivity results, not validated predictive effects. The parking predictor is derived from Census 2021 household housing/car margins and national EHS 2024 provision; actual local parking use is unobserved. It must not be described as measured garaging. Uncertainty Decile intervals are exact Poisson (Garwood), conditional on estimated mileage. They omit geographic heterogeneity and measurement error. Rate-ratio intervals additionally resample whole authorities within fixed extreme bins, 2,000 times, seed 20261010. All years for a sampled authority travel together. No interval accounts for unmeasured confounding, traffic estimation error or parking model error. Recorded parked involvement is a secondary endpoint. It is not proof that a parked vehicle was struck. Unknown coding is retained and bounded separately. STATS19 omits damage-only collisions. All-injury numerators include non-motor-only collisions; miles refer to all motor travel. Next gate Validate the parking measure, match residential-road exposure, and test incremental predictive value with confounder controls and geographic and time holdouts. Annual descriptive consistency is not a held-out test.