Label-free evaluation runs on a proxy: ask the same question several ways and treat agreement as evidence of correctness. Across 23,170 perturbed financial questions and two frontier models the proxy is real — an answer below the agreement threshold is about eight times more likely to be wrong — and it still waves 60.23% of all wrong answers through as unanimous. In 77.7% of those cases every rephrasing returned the identical wrong answer, so there was nothing to disagree about. Raising the threshold flattens out at 0.6, and asking a second model family produces three and a half times more shared errors than independence would predict.