pub const CALIBRATION_PRIOR_SAMPLES: f64 = 20.0;Expand description
Prior-strength constant for the empirical-Bayes confidence shrinkage.
Calibrated confidence = raw + w·(hit_rate − raw) with w = n/(n + k). With n resolved claims and k = 20, a base rate estimated from 20 samples carries half the weight of the raw confidence; fewer samples → weaker shrinkage, more → stronger. Exposed in the calibration report so the effective weight is always visible.