Brier score and Expected Calibration Error (ECE): lower is better. Curves closer to y = x are better calibrated.
Larger markers represent more predictions in that score bin. Each prediction scores one failure mode on one trace. Select a model or hover a point to inspect its scores and count.