Every six hours the Open-Meteo forecast for the observatory is logged; about a week later the ERA5 reanalysis for the same hours arrives and the two are compared, night hours only (18:00–05:00 UTC), grouped by how far ahead the forecast was made. This is what the planner's weather mark rests on — so here is how much it can be trusted, measured rather than assumed.
Logging since 2026-10-10 · 78 forecast-hours logged · 0 already compared with ERA5. The first comparisons arrive about a week after the first log run.
ERA5 is itself a model reconstruction, not a cloud sensor on the ridge — so this measures forecast-versus-reanalysis, the best public yardstick there is for this spot, not forecast-versus-sky. A persistent bias here would be folded into the planner's cloud factor once there are enough nights to trust it; until then the numbers are shown and nothing is corrected.
curl -sS https://devops.majbase.com/rozhen-planner/forecast-accuracy-apibuckets[] by lead time (n, cloud_mae, cloud_bias, overcast_agreement, rh_mae, wind_mae, temp_mae), recent[] per night, since, hours_logged, hours_compared. 503 when the log is unreachable.