FROST PHASE 6 — INCREMENTAL AUTHORITY SCREEN
Run from the same source checkout, with the immutable frost v1, phase4-v1,
phase5-v1 inputs and original STATS19 collision archive present:
  python scripts/build-frost-phase6.py
  python tests/frost-phase6.test.py
Requirements: numpy and scipy. Exact input hashes and runtime versions are supplied.
The bundle contains this phase's evidence, tests, code and the phase5 model engine;
it does not duplicate all original raw inputs. Keep the repository folder layout.

M0 is the frozen phase5 baseline. M1–M3 retain its mean structure/family and
refit coefficients only on 2021–2023. Development folds, controls and exposure
are matched. 2024 traffic and calendar terms carry 2023 values. Knots, centring
and scaling use training data only. No 2024 recalibration or 2025 modelling.
The baseline specification was previously selected on development data, so this
is a conditional extension comparison, not jointly nested baseline selection.

2024 and descriptive 2025 patterns were inspected previously. The baseline's
scientific gate is open. This authority screen does not validate LSOA deciles,
future weather prediction, home-postcode risk, insurance claims or causation.
Annual anomaly analysis is unavailable until a matched realised frost series is
acquired. Frost/ice surface counts are a distinct secondary outcome, not weather
exposure. Winter burden uses annual miles, never annual miles divided by 365.
