ENGLAND & WALES — ORIGINAL PHASE 5: INDIVIDUAL FEATURE EVIDENCE
Release ew-phase5-v1 | 13 October 2026 | Retrospective research only

WHAT THIS RELEASE DOES
Tests 38 nationally banded features: 28 new candidates and 10 functional-form
diagnostics for variables already in the frozen Phase4 comparator. Keeps all
62 candidates visible, including six unbanded diagnostic-only measures and 18
deferred measures. It does not create new bands, prove the paper, construct a
joint index or establish insurer-claims value. National reference population:
35,672 LSOA2021 areas, England and Wales. The 2022–25 panel retains 142,688 rows.

Read specification.json, frozen before fitting. Every feature uses its exact
existing values/bands. Missing or withheld data never become zero. LSOAs excluded
from a model remain in the national lookup. The richer Poisson baseline selected
in Phase4 is refitted on the same eligible cohort within every comparison.
All eligible zero-collision cells remain. Outcome: distinct motor-involved
police-recorded injury collisions. Offset: log(vehicle miles / 100000).

DEVELOPMENT AND INFERENCE
Only 2022–23 determines decisions. Train2022→predict2023; separately leave each
of nine English regions and Wales out across development. Unseen regions receive
the mean fitted log-region effect as in Phase4. Freeze the development fit for
2024/25 diagnostic predictions, carrying the last fitted time effect forward.
Both later periods have been inspected previously; neither is confirmation.

Categorical frozen bands are the primary test and govern retention. Compare
continuous linear and restricted cubic spline bases secondarily. Spline knots
are training-value quantiles, not new feature deciles. Baseline columns are
preserved; dependent added directions are removed. A zero-incremental-df linear
test is already represented. For sparse features, Z0 stays separate from positive
P1–P10. Empty national bands and absent training bands are explicit. A categorical
test row with an unseen training band receives the baseline forecast, is flagged
unsupported and blocks retention; this fallback is not an estimated neutral effect.

One categorical omnibus cluster-Wald p per feature; Benjamini–Hochberg FDR within
the 28-candidate primary family and separately the 10 baseline-form diagnostics.
No minimum p over representations, years or groups. Annual/regional/urban effect
patterns are exploratory. Coefficient intervals use a finite-sample corrected
cluster sandwich; Poisson overdispersion is not ignored via naive standard errors.
Authority blocks permanently join Cumbria successors and any authorities connected
by the actual inherited insolvency source reporting map. The cluster-map file
records all assignments. Financial source-area counts are also reported per band.

RESEARCH DISPOSITIONS
Retained for joint testing requires categorical q<=.05, mean temporal/spatial
deviance gain>=1%, both gains positive, >=7/10 geographic gains positive, both
conditional gain intervals above zero, no unsupported held-out bands, and positive
temporal gains under benchmark/length mileage and the matched counted-route scope.
Uncertain covers incomplete gates with some positive primary held-out gain or
q<=.05. Unsupported means neither primary held-out gain is positive and q>.05;
it is evidence against this tested incremental form, not proof of no association.
Untested candidates have uncertain/not-tested status and are not in the FDR families.
Baseline-form retention is a form suggestion, never a new-feature discovery.

PERFORMANCE AND EXPOSURE
Gain = 1 - feature Poisson deviance / baseline Poisson deviance, identical rows.
Positive means improvement. Frequencies = SUM(collisions)/SUM(miles)*100000,
never the mean of LSOA rates. All-road mileage alternatives share one eligible
cohort; categorical models are refitted independently. Better-observed all-road
temporal groups are diagnostic subgroups. The counted-major-route model uses the
Phase4 matched numerator and physical annualised counted mileage, its own baseline
road mix, and unchanged candidate feature values/bands. This selected major-road
scope avoids modelled minor allocation but is not national exposure validation.

Intervals for predictive gains resample 1,000 whole-block fixed prediction totals,
seed20261012. They do NOT refit models/selection or estimate exposure uncertainty.
They do not meet E8's formal promotion bootstrap requirement. Weak-support flags
are <100 LSOA-year rows, <100 collisions, or <10 independent blocks. Reference bands
can differ in sparse regional cohorts; read referenceBand before comparing effects.
For a variable already in the baseline, RR is its incremental band correction,
not its total feature effect. Continuous curves similarly show added contributions.
Two extremely sparse regional heat bands contain no collisions. Their finite
numerical log-rate approximations are not finite MLEs: adjusted effects/intervals
are withheld as not estimable. Their support and fitting records stay visible.
This reporting audit does not change any primary test or research disposition.

REMAINING LIMITS
Phase3 mileage is provisional. Regional baseline calibration failures and residual
spatial dependence remain. Mixed-vintage Census, air, insolvency, climate normal
and April2026 road context cannot be called information available at a 2022 forecast
origin. 2025 financial/air editions are retrospective context for earlier outcomes.
Climatology is not annual post-pandemic weather. All-road length-share controls
include restricted roads while mileage is access-screened. Terrain and vehicle
mix remain deferred. No causal, policyholder-residence, all-claims or pricing-factor
interpretation. Joint tests, shrinkage, earlier-collision comparator, refit-bootstrap
uncertainty, credible exposure, better calibration and untouched/claims evidence
remain for later phases. A statistically strong feature can remain uncertain.

FILES AND REPRODUCTION
register.json and feature-register.csv cover every candidate. features/<id>.json
is each evidence card. models/<id>.json.gz holds every fitted design, coefficient
and available covariance. blocks/<id>.csv.gz holds paired validation block totals.
Do not sum across repeated models, scenarios or samples. Aggregate exports cover
raw frequencies, adjusted bands, subgroup calibration and continuous curves.

Extract reproduce-ew-phase5-v1.zip into an empty directory. Use Python3.12 and
install pinned requirements.txt in a virtual environment. Run:
  python scripts/reproduce-ew-phase5.py
The runner sets OPENBLAS_NUM_THREADS=2 and OMP_NUM_THREADS=2. It refits every model
from included exact source inputs and compares every scientific output SHA256.
No network or GIS re-acquisition is required. Allow several minutes and several GB
of memory. Exact floating-point bytes can depend on BLAS/library versions; inspect
any mismatch rather than silently accepting it. The baseline script is included
for its frozen design; Phase4 is not rerun or edited. Prior published evidence is
pinned in frozen-prior-files.json. Source/output manifests accompany the release.

Underlying ONS/DfT/STATS19 and other public sources retain their recorded licences
and attribution. See E8 feature dictionary and the upstream source registers.
Ordnance Survey OpenData attribution applies to derived road context. There are
no insurer claims or personal policyholder records in this release.
