Your assistant answers ‘Alexander Fleming’ to ‘which sc...

In this scenario, the assistant is not hallucinating because the retrieved passages actually support the answer provided. Instead, this is a case of answer drift where the system is answering a different question than the one intended by the gold annotation. Because the retrieved content is technically grounded, standard groundedness checks will not trigger. The failure lies in the metric itself, as the gold annotation was likely written for a surface query that does not align with the actual information retrieved and processed by the model.