FROST DECILES AND COLLISION FREQUENCY — PHASES 1–3, VERSION 1

Implemented preparation, not a predictive finding.

England: 33,755 LSOA2021 areas. The primary public-source baseline uses
Met Office HadUK-Grid v1.1.0.0 2 km annual air-frost normals for 1991–2020.
OS Open Roads April 2026 supplies static mapped, non-fictitious road-length
weights. The separate land-area sensitivity has its own fixed England deciles.
The source period precedes outcomes; the road snapshot does not. This is a
retrospective geographical study, not an exact historical-availability backtest.
The preferred latest 1 km CEDA upgrade still requires source access.

WHAT IS READY
Protocol, real frost lookup, fixed deciles, compact weights, collision audit,
765 authority-year rows, 682 frost-screening eligible rows,
131 authorities eligible in every study year, and 168,775 LSOA-year
collision inventory rows. Positive-mileage zero collisions remain zero.

WHAT REMAINS UNAVAILABLE
No validated LSOA motor-mile denominator, no major-road LSOA rates, no frost
decile rates, no fitted model, no incremental prediction result, no causal or
insurance-claims interpretation. DfT MRDB links are endpoint chords. A physical
centreline, validated collision/link matching and credible unsampled minor-road
exposure are still required for the original all-road LSOA frequency question.

REPRODUCING FROM A SOURCE CHECKOUT
Install the packages pinned in scripts/frost-requirements.txt in an isolated
Python environment. Source ZIPs are large and are intentionally not copied
into the compact download. Download the exact public source editions recorded
in source-register.json and verify hashes before use. Keep the existing frozen
STATS19 and DfT archives at their repository-relative source paths.

1. python scripts/prepare-frost-collisions.py --events-dir INPUT_DIR
2. python scripts/build-frost-features.py --input-dir INPUT_DIR
3. python scripts/audit-frost-major-roads.py
4. python scripts/build-frost-panels.py INPUT_DIR
5. python scripts/package-frost-study.py INPUT_DIR
6. python tests/frost-study.test.py

INPUT_DIR contains the acquired Met Office JSON, ONS source ZIP and shapefile,
their acquisition manifests, and major_roads/os-open-roads-april2026.zip plus
its manifest. The feature builder verifies raw hashes and extracted boundary
members. See its --help for the exact contracts. No CEDA login is needed for
the public 2 km baseline. Never overwrite or silently relax protocol.json.

COMPACT REPLAY FROM THE DOWNLOAD
Unzip retaining paths. The compact archive includes the numerical inputs and
geography context needed to check frost calculations and rebuild the prepared
panels. It does not contain 800+ MB of original road and boundary ZIPs.

python scripts/build-frost-features.py --from-weights public/research/frost/v1
python scripts/build-frost-panels.py . --panels-only
python tests/frost-study.test.py

The first command recalculates and verifies numerical scores and cutpoints
from weights, using the feature CSV only for area identifiers and metadata.
The second rebuilds authority joins and LSOA inventories from the already
audited collision assignments. Full source-geometry reproduction uses the
large originals and command sequence above. Source hashes, data dictionary,
feature audit, spatial audit and major-road audit explain all exclusions.

ASSESSMENT RESERVED FOR LATER PHASES
Development 2021–2023; validation 2024; first frozen frost-model assessment 2025.
Other project studies have used 2025 outcomes; this is not an untouched future
year. Whole geographical units remain together. Keep continuous frost and
fixed bands as alternative candidates; do not promote a claim from a map.

ATTRIBUTION
Contains Met Office data licensed under the Open Government Licence v3.0.
Contains OS data © Crown copyright and database right 2026. ONS and DfT
data and licences are recorded in the source register.