Calibration Surface — Predicted vs Actual Win Rate

A perfectly calibrated model traces a diagonal ridge: predicted 70% → actual 70%, regardless of year. Bumps above the diagonal = model is under-confident; dips below = over-confident. Gaps = no fights landed in that (predicted prob × year) cell, or the cell was filtered out by the min-n toggle. Sparse cells (≤10 fights) carry high binomial noise — toggle Min n=20 or Min n=40 to drop them. Source: api/calibration-surface.json · 8,000+ walk-forward predictions