ENGLAND & WALES — F7 ADDITIONAL-FEATURE COLLISION SCREEN
Release ew-f7-v1 | 13 October 2026 | Retrospective development evidence

RESULT
Two source-admitted features tested; nine candidates explicitly untested.
Public transport access-site density: retained for F8 joint testing, with
3.56% temporal and 3.43% pooled geographic deviance improvement (mean 3.49%),
9/10 regions positive. Road-weighted heavy-rain days: uncertain; 0.24% temporal
and -0.69% geographic gain (mean -0.23%); 5/10 regions positive; FDR q=0.0745.
The same-observation sensitivity gives the same conclusion. These are relative
Poisson-deviance gains, not percentage reductions in accidents or claim costs.

The expanded ranking has 30 novel candidates (28 original plus two F7), with
ten original baseline-form diagnostics recorded separately. Access-site density
ranks 12/30 and heavy rain 24/30 on the equal-weight primary mean. Original
results, q-values and test-family definitions are unchanged. Different feature
cohorts mean that small cross-feature ranking differences are descriptive.
There are now 16 retained single-feature candidates across the two releases.
F8 joint testing, joint indices and F9 report refresh are not implemented here.

NATIONAL BANDS AND COHORTS
All 35,672 England and Wales LSOA2021 areas retain the frozen F6 national values,
bands and missingness. No local/fold/cohort deciles or substituted values are
created. Eligible zero-collision observations stay in the models. The full panel
has 142,688 LSOA-years (2022–25). The baseline requires valid positive mileage
under all three national exposure scenarios and complete baseline controls.
Feature-specific eligible counts and exclusions appear in every card and the
cohort-coverage export. National coverage is 35,646 access-site LSOAs and 35,365
heavy-rain LSOAs; all unavailable areas remain in the upstream national lookup.
The common development cohort comprises 32,789 LSOAs / 65,578 observations.
The two single-feature models and their baselines are independently refitted
on this same intersection; no joint feature model or additional FDR is implied.

FROZEN METHODS
See specification.json and the approved STATS19 Additional Features Phased Plan.
Baseline: frozen Phase4 rich_poisson implementation, refit on exactly the same
rows as the candidate in every fold. Outcome: distinct motor-involved
police-recorded injury collisions at collision location. Offset log(miles/100000).
2022 fit -> 2023 prediction; ten leave-one-region-out folds across 2022–2023
(nine English regions and Wales), pooled out-of-fold geographic deviance.
Higher national band means more of the feature, not automatically more risk.
Categorical bands are the sole primary representation. Continuous linear and
restricted cubic spline forms are secondary: log1p access-site density, identity
heavy-rain days; training scaling and knots refit inside each fold.

Rank = (temporal relative deviance gain + pooled geographic relative deviance
gain) / 2. Gain = 1 - candidate deviance / same-cohort baseline deviance.
Raw frequency = SUM(collisions) / SUM(miles) * 100000, not mean local rates.
O/E is observed / expected collisions. Geographic calibration remains imperfect:
the access-site primary model's pooled geographic O/E is about 1.134. Better
relative deviance does not itself resolve baseline calibration or exposure bias.

One final-development categorical omnibus cluster-Wald test per primary
candidate. Benjamini–Hochberg over all eleven candidates; the nine unavailable
primary tests contribute conservative p=1 placeholders. Their p/q=1 is not an
observed null finding. Original 28-candidate and ten baseline-diagnostic families
stay unchanged. Combined 39-candidate sensitivity is deferred until all eleven
additions are available; no completed combined-family adjustment is claimed.

RETENTION GATES
q<=.05; mean primary gain>=1%; positive temporal AND geographic gains; >=7/10
positive regional folds; both conditional gain-interval lower bounds positive;
no unsupported held-out categorical bands; positive temporal gain under traffic-
group benchmark, road-length exposure and the matched counted-route scope.
Uncertain means an estimable feature misses at least one gate while showing
some positive primary performance or q<=.05. Unsupported means neither primary
gain is positive and q>.05. Untested is kept separate from these dispositions.
Untested candidates: schools, claimant rate/change, mean gradient/steep share,
heat >=25C / >=30C days, wet-day rainfall intensity and snow-lying days.
F6 measurement/coverage decisions cannot be overridden using collisions.

UNCERTAINTY, SUPPORT AND SENSITIVITIES
Finite-sample corrected cluster sandwich coefficient intervals. Permanent
Phase5 blocks union authorities sharing inherited insolvency source areas,
including Cumbria successors. Prediction-gain intervals use 1,000 paired block
resamples of fixed prediction totals (seed20261012). They omit fitting,
selection and exposure uncertainty, and do NOT satisfy the formal E8 promotion
gate. Subgroup effects are exploratory. Weak support: <100 rows OR <100
collisions OR <10 independent blocks. Empty/zero-event bands are not given
finite effect estimates. Unseen held-out bands fall back to baseline prediction,
are flagged unsupported and prevent retention. Read referenceBand in each fit.

Alternative-mileage models share the same feature-specific eligible rows and
are separately refit. Counted routes use matched route-scope collisions and
physical counted mileage, with route-specific baseline road shares. This is a
restricted major-road sensitivity, not representative all-road validation.
Models, cluster blocks, annual/region/urban effects, raw frequencies and held-out
subgroup calibration are retained, including negative and null results.

SOURCE VINTAGES AND REMAINING LIMITS
Access-site density is a 13 October 2026 NaPTAN snapshot near April2026 mapped
roads: physical access sites, not service frequency, ridership, walking or cycling
flow. Heavy rain is an interpolated 1991–2020 climate normal: annual count of
days >=10mm rain, not actual annual post-pandemic weather. All valid heavy-rain
values are estimated grid-normal context; there is no observed-only counterpart
on which to run a separate measured-weather test. Exact F6 source metadata is
preserved inside cards as frozenSourceMetadata; current descriptive notes clarify
the F7 evidence status. The immutable F6 release remains the data provenance.

2024 and 2025 were already inspected and are diagnostics only. No untouched
confirmation, as-of-2022 forecast, causal effect, driver-residence effect, insurer-
claims validation or pricing factor is established. Mileage allocation remains
provisional. Joint incremental value, overlap, shrinkage, earlier-collision
comparators, improved calibration and refit/selection uncertainty remain for F8
and subsequent validation. The original paper's hypothesis is not proved here.

REPRODUCE
Extract reproduce-ew-f7-v1.zip into an empty directory. Use Python3.12 and the
pinned requirements.txt (same numerical environment as Phase5). Run:
  python scripts/reproduce-ew-f7.py
The runner checks all input hashes, sets OPENBLAS_NUM_THREADS=2 and
OMP_NUM_THREADS=2, fits all 134 saved models (112 primary/secondary/subgroup/
sensitivity fits plus 22 common-cohort fits), and compares every numerical
output SHA256. No network or source re-acquisition is needed. Exact bytes may
depend on BLAS/platform versions; investigate every mismatch. The original
Phase4/5 code is included unchanged for its frozen design, not rerun as a new
release. New implementation uses build-ew-f7.py. All 736 prior public scientific
files are pinned in frozen-prior-files.json and checked unchanged before release.

FILES
register.json / feature-register.csv: all eleven candidates and gate decisions.
expanded-ranking.json/csv: 40 tested features; 30 ranked novel candidates and ten
baseline form checks. Original family q-values remain original-family values.
features/<id>.json: complete evidence cards; models/*.json.gz: fit parameters;
blocks/*.csv.gz: paired fixed-prediction uncertainty totals. Do not sum across
repeated samples, model forms or scenarios. Common-cohort files include exact
membership, metrics, models and blocks. Aggregate exports contain all band
relativities, raw frequencies, subgroup calibration and continuous curves.

Public data retain the licences/attributions recorded in upstream source
registers, including DfT/STATS19, ONS, Met Office and Ordnance Survey derived
road context. No insurer claims or policyholder records are used.
