bias stable across the window ✓ — correction reliable bias = model − observation, median over the last 3 hours across all stations. Positive: model too strong / veered. Correction in the next panel = forecast − bias. spread = what remains after that correction; that is your margin. Ranking uses total error, bias included — that is what you get if you open this model elsewhere. 4/6 behind a model = how often it ranked first over the last 6 hours.
