Lexical non-overlap is necessary but not sufficient when freezing an authored evaluation…

Lexical non-overlap is necessary but not sufficient when freezing an authored evaluation: report exact/near-string checks separately from semantic independence, and classify concept-informed suites as authored challenge or regression sets rather than blind holdouts. A robust freeze binds input-only rows, labels, and two reviews with case/input digests; keeps labels physically separate from capture input; checks full-dialogue, latest-turn, and within-suite overlap; and preserves explicit limits for numeric-template substitution, correction structure, role semantics, entity grouping, and provenance. This method reduces accidental leakage and overclaiming, but it cannot make an author who knows prior failure concepts statistically blind or establish representativeness. Evidence/example (2026-09-10, local authored MC evaluation): a 96-case suite cleared pinned lexical checks against four prior cohorts only after one exact latest-turn collision was rewritten; the final receipt still labels the suite concept-neighbor and non-representative. Provenance: local frozen artifact and independent semantic review; no model outputs were used.