Verification
No provider publishes how well it actually performs. This measures them against real observations and shows the result.
Every provider WayPilot queries is asked, on an hourly cycle, what it expects at a fixed set of airports worldwide, for +15/+30/+60 minutes ahead. Once that time has passed, the call is checked against the nearest METAR observation from aviationweather.gov and scored. POD is how much real rain a provider caught; FAR is how many of its rain calls were false alarms; CSI folds both into one number; Brier scores providers that publish an actual probability rather than a hard yes or no. POD, FAR and CSI are only shown once a bucket has at least 20 matched predictions AND at least 10 observed rain events. Without observed rain there is nothing to catch, so FAR collapses to 100% and CSI to zero for anyone who forecast rain at all, which looks like a verdict and is really an absence of weather.
No predictions have resolved yet. This is not a verdict, it is silence: the hourly cron has to run a few times before any forecast has had a chance to come true against real weather. Check back in a few hours.
Reference stations, scoring formulas and the raw prediction/observation log live in lib/verify/ and data/observations.jsonl — this page renders data/scores.json, nothing more.