ENGLAND & WALES — PHASE4 GEOGRAPHIC AND ROAD BASELINE
Release ew-phase4-v1 | 13 October 2026 | Retrospective research only

This release implements original Phase 4. It does not test the 38 frozen feature
bands individually, promote risk indices, prove the paper or validate insurance
claims. All 35,672 national LSOAs and 142,688 annual records remain accounted for.
The 18 deferred candidates remain deferred. See specification.json before use.

RESULT
The selected development comparator is rich_poisson: a Poisson count mean with
log(miles/100000) offset, geographic/year effects, density/age/road mix,
outcome-free exposure quality, nonlinear density terms and richer road layout.
Poisson overdispersion is material. NB2 estimates and clustered coefficient
uncertainty are published; NB2 did not improve the prespecified combined
out-of-time/geographic prediction criterion. Poisson selection does not imply
Poisson variance is correct. Do not use naive Poisson standard errors.

Development uses 2022–23 only. Fit 2022→predict 2023 and leave one of nine English
regions or Wales out across both development years. Latest fitted year effects
are carried forward for 2024/25 without using their outcomes for recalibration.
2024 is a selection-year diagnostic;2025 was previously inspected and is
exploratory. All 2024/25 prediction gains are retrospective, not confirmation.

A separate history comparator uses (earlier observed+10)/(earlier expected+10)
times the baseline mean. Frozen history uses eligible 2022–23 rows. The separate
2025 rolling-history comparator also uses 2024. These are local-history forecasts,
not cold-start spatial validation. Missing history receives multiplier 1.

All three mileage scenarios use exactly the same rows and selected structure;
the coefficients are refitted to each scenario. Road mix is the frozen static
length measure and includes restricted-access roads, unlike the access-screened
mileage network. Density and road attributes also enter exposure estimation.
No terrain or vehicle-mix control is invented for deferred fields.

CALIBRATION AND LIMITS
Material regional bias remains. Wales O/E is about 0.890 in 2024 and0.891 in 2025;
London and North East also fail the 0.90–1.10 diagnostic tolerance in 2025.
Out-of-region transportability and spatial residual dependence remain concerns.
Phase 3 contains zero strict-exposure rows and zero validated LSOA frequencies.
Use this as a conditional research comparator. Phase 5 can investigate incremental
value but cannot promote commercial indices while these gates remain open.

Coefficient intervals cluster by UTLA, merging both Cumbria successors because
they share the 2022 control. Performance intervals resample whole clusters 1,000
times with seed 20261012, conditional on fixed fitted predictions. They omit
model/selection refitting and mileage uncertainty. They do NOT satisfy the E8
promotion bootstrap gate. Cross-authority residual dependence is reported.
Prediction calibration groups are diagnostic rank groups within a result sample;
they are NOT new national feature deciles and never overwrite the frozen bands.

COUNTED MAJOR-ROUTE SENSITIVITY
Counted AADF, accepted routing and no route overlap are required. Only distinct
motor collisions with unique compatible 30m road matches AND membership in the
same clipped LSOA/route section enter its numerator. Exposure uses annual AADF
multiplied by actual accepted physical route length and calendar days. Zero
collision cells remain. Separate restricted-scope road mix and model refit are
used; this is not an all-road versus major-road risk comparison. It avoids minor
allocation but still depends on annualised counts,2026 geometry and imperfect
collision matching. See counted-route-summary.json and collision-membership file.

REPRODUCE
1. Extract reproduce-ew-phase4-v1.zip into an empty directory.
2. Use Python 3.12 and a virtual environment. Install requirements.txt.
3. Run: python scripts/reproduce-ew-phase4.py
The runner sets OPENBLAS_NUM_THREADS=2 and OMP_NUM_THREADS=2, refits the complete
national analysis and counted-route sensitivity, then compares every scientific
output against model-output-checksums.json. Allow several minutes and several GB
of RAM. The kit includes all exact upstream bytes needed; no raw GIS download or
network service is required. Floating-point libraries/BLAS can affect exact byte
matching outside the pinned environment. Never ignore a mismatch silently.

INPUT AND OUTPUT CONTRACT
input-checksums.json pins every source used in fitting or counted-route matching.
frozen-prior-files.json audits preservation of all prior public scientific files.
manifest.json pins delivered output bytes (excluding itself/release verification).
model-output-checksums.json lists every numerical reconstruction output.
models.json stores selected coefficients, column centres/scales, region contrast
levels, year levels, density knots and dropped rank-deficient columns. See code
for scoring. A predicted value is a conditional area-of-collision injury count,
not an insurer score, policyholder-residence risk or a causal effect.

The full validation predictions are long-form: one row per LSOA/year/model/
scenario/sample. NEVER sum across repeated models, scenarios or samples.
National predictions contain in-sample fitted 2022–23 values and frozen out-of-time
2024/25 values; the sample column distinguishes them. Unmodelled rows remain blank.
O/E is SUM(observed)/SUM(expected). Frequency is SUM(collisions)/SUM(miles)*100000,
never an average of small-area rates. Zero-collision rows with positive exposure
are included; invalid denominators never receive a rate.

SOURCE LICENCES
Underlying ONS/DfT/STATS19 records retain the upstream source-register licences
and attribution. UK Government/ONS public records are used under their recorded
Open Government Licence terms; Ordnance Survey OpenData attribution applies to
derived road context. See pinned Phase 3 design and E8 feature dictionary/source
registers. No insurer claims or personal policyholder records are included.
