Semantic similarity metrics spend under 4% of their range on a legal meaning reversal.

LexFlip attacks the standard validity check for meaning-preservation metrics, which requires an identical pair to score highest and an unrelated pair lowest, and therefore passes any monotone function of token overlap (arXiv 2609.05296). The dataset releases 373 minimal perturbations of Quebec statutory French that reverse legal force while preserving 0.93 of tokens. Seven embedding and BERTScore metrics spend 0.022 to 0.039 of their identical-to-unrelated range on such an edit, against 0.670 for bidirectional NLI, the one family the conventional check disqualifies. On FrJudge, a bare length feature outscores every semantic metric.