Metamorphic Evaluation of Language Models: Robustness Under Semantics-Preserving Rewrites

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Paired, program checked protocol for measuring whether a correct answer survives a change in surface form and what that change costs in generated tokens. This study involves sixteen models across three providers and three family ladders.