Eval results

Each run sends the labelled question set through the real pipeline. Retrieval is scored against the documents each answer needs; answers are checked for required facts and correct refusals, and optionally graded by a separate judge model for correctness and grounding in the retrieved text.