A prediction task from NOAA's tide gauges

122 NOS stations, 2006-2025, 2,440 station-years of verified water levels: will a station's maximum observed water level exceed its own published minor flood threshold tomorrow? NOAA publishes the observations, the verified maxima and the annual counts, so access is not the contribution - the frozen panel, the observations-only feature contract and the leakage gate are.

The release carries two levels, each a complete task instance. The temporal level is the ordinary panel split. The station-disjoint level holds 41 stations out of training entirely, and it is the point of the dataset: it asks whether a model learned something transferable rather than one station's local habits. It scores the same on stations it has never seen.

The numbers are published with their floor. The baseline reaches 0.8633 (temporal) and 0.8690 (station-disjoint); a single unlearned feature - yesterday's maximum against the station's own threshold - already reaches about 0.83 and 0.85. The headroom is narrow, and a reviewer should see that here rather than discover it. Only about 2% of station-days are positive, so a constant answer agrees with the label 96% of the time: score AUC, never accuracy.