calibration/

Post-processors that turn raw classifier scores into calibrated probabilities:

• PlattCalibrator: parametric sigmoid fit; the default for small

calibration sets and margin-like scores.

• IsotonicCalibrator: non-parametric monotone step function via PAVA;

more expressive, needs more calibration data.

• TemperatureScaler: single-parameter multiclass scaling that preserves

accuracy exactly.

• VectorScaler: per-class scale and bias for class-dependent

miscalibration.