Held-out accuracy per prediction model. Pass when Aito's accuracy beats the baseline by ≥ 10 pp. Fail means Aito is honestly telling you it doesn't know — even when the raw accuracy is high.