Calculation
What we compute after a window closes
A scenario is scored only when its horizon ends or a fact resolves it. Live charts on /statistika use those closed rows. We do not backfill a class Brier to look better.
Brier
Average of (forecast probability − outcome)². Lower is better. A coin on a binary event is 0.25. Mixed books can sit near that line.
ROC-AUC
Separation of classes. 0.5 is chance. If a short window has only one observed class, AUC is not interpretable — we say so instead of inventing a number.
Model A 0.6339
Named risk-model AUC from the published dual-model audit. Product anchor, not a class table.