Performance
All numbers are out of the archive: every 2025–2026 frost-season night, forecast, calibrated, and compared with what the station actually measured. No cherry-picking.
The scorecard
Alert rule: protect when P(frost) ≥ 10% (adjustable):
A false alarm costs one night of protection; a miss can cost the crop — at α = 0.05 (cheaper protection) all three frost nights alert at every lead 1–3.
Error vs lead time
Typical error is ≈0.8 °C the night before and grows smoothly to ≈1.6 °C two weeks out.
Honest probabilities
Flat bars at 0.10 = the forecast uncertainty matches reality; a "30% frost risk" means 30%.
The frost nights themselves
The calibration corrects uncertainty from weather forecast models and improve the prediction of frost events.
High-confidence false alarms (P ≥ 50%, 1–3 nights ahead), whole archive:
0 .One site, ~2 seasons, 3 frost nights so far — event metrics are provisional; temperature metrics use all ~200 nights. Full methods & math: the back-analysis report in the repo. Static snapshot; production rebuilds every forecast cycle.