It never stops watching.

Every week of flu season, Prime Radiant forecasts hospital admissions for every US state — 23 probability levels at a time — built to file them to the CDC FluSight hub, the ensemble public-health teams use (registration in review; first live submission this season). This is the 2025–26 season, replayed: real reported data, and the forecasts the model would have made at the time — computed only from data that existed each week.

How to read this honestly. This is a replay, not a live feed: reported admissions are the hub's own truth data, and each forecast was computed from only the data that existed that week — the model trains on the hub's git history as it stood at each forecast date, never on what came after. A flag one week ahead of the crossing is what 1-to-2-week-horizon forecasting of a fast ramp honestly buys; the 23 flags that land on the crossing week itself still precede the report that confirms them, because reports lag the week they describe — and the five states flagged a week behind are printed in the score, not hidden. Known debt, also shown: the displayed ensemble's uncertainty bands are too narrow (50% intervals cover ~37–46% — 46% in this season), and a stricter version of this rule using the conservative quartile is never early and usually late. The fix (season bagging) is parked and public in the repo's model card.