ParaRecover scores whether an agent notices a mid-run failure, not whether it finished.

A 10,626-instance benchmark built on a 14-type error taxonomy, scoring structural integrity, diagnostic reasoning and recovery strategy during execution rather than at the end. Across ten-plus mainstream models, multi-turn error propagation and implicit tool-use failures remain unsolved, and the rubric itself doubles as a supervision signal. This is the eval most agent teams don't have, because task success hides everything that went wrong on the way to an accidentally correct answer.