Persisting calibrators

How-to; what is stored, why JSON and never pickle, and the compatibility promise live in the Serialization concepts chapter.

cal = BetaCalibrator().fit(scores_cal, y_cal)

cal.to_json("calibrator.json")                  # human-readable, versioned
loaded = BetaCalibrator.from_json("calibrator.json")
# ...or without knowing the class:
from probcal import BaseCalibrator
loaded = BaseCalibrator.from_dict(json.loads(text))   # registry dispatch

cal.fingerprint()                # sha-256 provenance id (identical fits match)
cal.fit_meta_["data_fingerprint"]  # sha-256 of the sorted training triple

Objects with external parts keep them as references, never blobs: CalibratedModel.from_dict(d, model=...) reattaches the base model; CalibratedScorecard.from_dict(d, scorecard=...) verifies the scorecard table's fingerprint before attaching. The monitor serializes its whole past (CalibrationMonitor.from_json resumes bit-for-bit).

Every 0.x release reads schema 1 — enforced by committed golden files in CI, not promised on trust.