MatchCalibration

Measuring how well a match configuration performs against known answers, and tuning it on evidence.

📄️Evaluate a profile against labelled pairs

Scores the profile against pairs a steward has judged and reports precision and recall — how many of its merges were right, and how many real duplicates it missed. This is the only way to know a profile is good rather than merely plausible. Requires `match.profile.evaluate`.

📄️List evaluations

Returns past evaluations, so a configuration change can be judged by whether accuracy actually moved. Requires `match.profile.evaluate`.

📄️Suggest thresholds

Recommends thresholds from the labelled set: the automatic-merge line is placed above every judged non-match, so that everything merged without review is right, and the review line is chosen for the best overall accuracy. The recommendation is returned for you to apply, never applied for you. Requires `match.profile.optimize-thresholds`.

📄️Calibrate attribute weights

Fits how much each kind of agreement is worth from the labelled pairs, rather than leaving them at hand-set guesses. A calibration that would make accuracy worse is reported and not applied. Requires `match.profile.calibrate-weights`.