Cloud forecast, tested.

Peirce skill is the observed-clear detection rate minus the probability of falsely detecting clear sky during an observed-cloudy hour. It measures discrimination after accounting for how often the sample was clear: +1 is perfect, 0 has no discrimination, and a negative value is worse than the corresponding constant decision. Heidke skill is shown as a secondary chance-corrected score. The observed clear-hour base rate sits beside every skill score, and samples above 85% clear are flagged as low-information. The two cohort classes are never averaged into a published headline. GoodSeeing's NASA climate pages measure usable nights, not hourly clear occurrence, so that different-granularity statistic is not substituted for an hourly base rate. This verifies cloud only—not the full observing score. Satellite masks can miss thin cirrus, smoke and haze, and they do not measure atmospheric steadiness. Forecasts come from MET Norway; independent masks come from NOAA GOES, EUMETSAT Meteosat and Himawari. Comparable results begin with the first full observing night on 25 August 2026. Observations never alter the forecast shown for that hour. Download the rolling contingency counts and metrics as CSV.

25 August 2026: the comparable series was reset to require exact UTC-hour matches, one astronomical-darkness boundary and version-qualified satellite masks; earlier display-resampled observations remain archived but do not enter public results.

3 September 2026: the publication counter was clarified as a completed-night snapshot. Incoming observations can increase during a day, but a night counts only after the UTC date closes and its required coverage passes.

4 September 2026: the publication metric was frozen before release: Peirce skill leads, observed base rate is adjacent, populated and dark-site cohorts are reported separately, and raw contingency counts are downloadable.