A mismatch between training (using ground-truth prefixes) and deployment (using model-generated prefixes) that causes errors to compound.