WeatherNext 3 — on the bench

Google DeepMind’s 64-member ensemble, two Pribilof cells, scored against what actually happened. Wind only — 10 m p10–p90 vs METAR and the buoys, with HRRR-AK in the next 48 hours. It is not a fog tool, and a 25 kt line is wind for the fleet, not seas. Just opening the flaps? What’s in the box.

Sandbox experiment. Not on the approach pages, not in the commentary, not advice. HRRR-AK still owns the next 48 hours on the weather pages.

Now vs the model

Loading…

The spread

Shaded band is p10–p90 — where eight in ten members sit. Gold dashes are 25 kt wind, not seas. Hover or tap a time.

St. Paul St. George HRRR (48 h) 25 kt

St. Paul — WeatherNext vs HRRR-AK

St. George — WeatherNext vs HRRR-AK

Days 3–15 — chance of >25 kt wind

p90 ≥ 25 kt means at least a tenth of the ensemble is at small-craft wind. Compared to the NWS PKZ766 nearshore wind lines, never the seas lines.

Track record

Observed 10 m wind inside the p10–p90 envelope is a hit. Within 5 kt of it is close. Percentages wait until n ≥ 20.