标签
This paper formalizes when benchmark contamination is detectable, deriving information-theoretic limits and proposing power-calibrated audits that distinguish a clean benchmark from a powerless detector. It reports two-sided empirical findings on calibration efficacy and validity gates.