Rewind-Bench: catching AI code fixes that only look correct
7% of an AI agent's "successful" code repairs stopped working the moment the data changed. Rewind-Bench re-runs every fix on data the model never saw, and counts the ones that were only pretending.
