Accuracy.

Perfect — prediction matches reality

Posterior — public calibration

Reading the chart.

When the model says 0.70, the underlying event should occur roughly seventy percent of the time. The closer the curve hugs the diagonal, the more our stated confidence matches the world.

Why it matters.

A miscalibrated model can post a healthy hit-rate while sizing bets at the wrong confidence and bleed bankroll silently. Candidate calibrators — isotonic, Beta, BBQ — are evaluated daily against resolved entries; one is promoted only after minimum-sample, holdout, and stability gates clear.

Hover a point.

Each bucket displays its predicted probability, the empirical hit-rate, and the sample size — the bigger the n, the more reliable the bucket.