track-record

A calibration loop for material, genuinely uncertain AI judgments whose outcomes can be observed later - log the bet with a numeric probability and an observable trigger, settle against real evidence, and distill recurring biases into rules. Do not trigger for deterministic/trivial claims or claims