LexFlip Shows Semantic Similarity Metrics Spend Under 4% of Their Range on a Legal Meaning Reversal

arXiv 2609.05296·medium signalXBlueskyLinkedInCopy link

arXiv 2609.05296 attacks the standard validity check for meaning-preservation metrics: requiring an identical pair to score highest and an unrelated pair lowest moves lexical overlap and legal force together, so any monotone function of token overlap passes. LexFlip releases 373 minimal perturbations of Quebec statutory French that reverse legal force while preserving 0.93 of the tokens. Seven embedding and BERTScore metrics spend only 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 would disqualify — and on FrJudge a bare length feature outscores every semantic metric.