The atmosphere's competing futures, counted member by member. Pick a pool — the ECMWF ensemble alone, the full-physics pool (ECMWF EPS + the HRES member + GFS + every GEFS member), or the A.I. pool (AIFS-ENS + AIFS + all of WeatherNext) — a region the clustering listens to, a lead or window, and how many scenarios to cut. Each cluster then shows what it would do: its 500 hPa pattern, its North-American temperature anomaly, its degree-day shift on the census population weights, and which stations move most.

Clusters re-form anonymously every run — but the storylines persist. Each row is one scenario tracked across successive inits by matching its z500 composite to the previous run’s (area-weighted pattern correlation, optimal one-to-one assignment). Watch the probability mass flow: a share climbing run-over-run is the ensemble committing to that future; a falling one is a scenario dying. ★ marks a birth; r is the pattern match against the previous run.

Build your own pool. Pick any set of shipped ensembles, a region the clustering listens to, a lead or window, and k up to 6 — computed on the spot from the same member fields, scored with the same 5-day accumulated degree days against your pool's own ensemble mean.

All pools, leads, regions and cuts at once — each scenario's CONUS temperature anomaly and degree days, and the Δ columns: what that scenario adds or subtracts relative to its pool's plain ensemble mean (the no-clustering baseline, shown as its own rows). Click any column to sort.