Verification ledger

Method & caveats

The observed bucket is this site's own scorer re-run with the buoy's sea substituted into the forecast's dominant partition, scaled by that partition's share of the total. The period substituted is the buoy's TOTAL-spectrum MEAN period (Tm), because Open-Meteo's swell_wave_period — what the scorer was tuned on — is also a mean period; on a windsea day Tm can sit under the scorer's hard period floor while the peak period is comfortably above it, which disqualifies the re-score rather than the forecast. Those rows are counted as rescore_disqualified, never silently dropped.

period_fallback_v1 changes exactly one thing about the strict re-score: when the buoy's TOTAL-spectrum MEAN period (Tm) falls under the spot's period floor while its PEAK period (Tp) does not, the substitution keeps the FORECAST's own partition period instead of substituting Tm. Those are windsea days — a long groundswell under a short local sea — where Tm is dragged under the floor by waves that are not the ones the break works on, so the strict re-score records observed = flat for a definitional reason rather than an observed one (live 2026-09-01: 757 of 759 period_below_min rows had Tp >= 6 s AND a forecast partition period >= 6 s). The forecast's period is used rather than Tp because Tp is the peak of a mixed spectrum and is biased high by several seconds against the mean period the scorer was tuned on, whereas the forecast period is unbiased against Tm (-0.6 s live); substituting a Tp would invent a better sea than the buoy saw. Keeping the forecast's own period means the variant tests HEIGHT and DIRECTION only on those rows, and says so.

46115 of the published verdicts read "flat" because the scorer disqualified the observed sea rather than scored it. Most of those are a period problem, not a forecast problem: the buoy publishes a total-spectrum mean period (Tm) which we substitute for the forecast swell period, and on a windsea day Tm is dragged below the scorer's hard floor while the actual peak swell period is fine. Those rows count as misses against us; we show the number rather than hide it.

12228 zero-hour nowcasts are held back. A 0 h "forecast" is scored from the same hour it is checked against, so it measures the model's agreement with the buoy, not our ability to see ahead — it stays a private control rather than a published hit rate.