How we grade the model: Brier scores, calibration, and honesty

Accuracy claims mean nothing without a grading rule that is fixed in advance. Ours is simple and public.

Winner accuracy counts a projection as correct when the side with the higher probability wins. It is intuitive but coarse.

Brier score measures the squared gap between the stated probability and the outcome. A model that says 60% should be wrong four times in ten.

Calibration buckets compare stated probabilities against observed frequencies. If our "70%" games win 70% of the time, the model is calibrated — even when individual games lose.

We never delete poor results, never hide losing forecasts, and never claim historical accuracy guarantees future performance.