Where evidence quality gets hard

These cases come from error analysis and challenge hardness, not invented drama. Compare expected labels with model predictions when receipts exist.