semantic_router.calibration

turn a raw router score into P(route is correct). monotonic, so ranking never changes.

Calibrator Objects

class Calibrator()

a fitted score -> probability map. build one with Calibrator.fit.

ece

def ece(probs: np.ndarray, correct: np.ndarray, bins: int = 10) -> float

expected calibration error: bin-weighted gap between stated confidence and observed accuracy.