What problem does it solve? A multi-lens AI review panel cannot judge its own quality from production records, since accepted and rejected counts measure adjudications rather than actual defects. This Skill evaluates the panel against fixtures whose defects are known in advance, producing honest catch-rate and false-positive measurements per lens. ## Core Features & Use Cases - Fixture-based evaluation: Runs the six review lenses against calibration cases with planted defects declared in manifests, then scores which defects each lens caught. - False-positive measurement on clean controls: Counts findings at minor severity or above only on control cases where nothing was planted, keeping the metric falsifiable. - Honest reporting discipline: Prints unmatched findings without judging them, exempts lenses via declared notControlFor rules, and refuses to be wired into CI or test suites. - Use Case: After changing a lens remit or swapping reviewer models, run the calibration to decide whether the six-lens panel still earns its cost before trusting it on production diffs. ## Quick Start Ask the AI to run the calibrate skill to evaluate the review panel against the calibration fixtures and report the worst catch rate and any false positives.
