Penalty-aware evaluation framework addresses RAG hallucinations

A new evaluation framework for RAG systems uses asymmetric scoring and knowledge-gap canaries to penalize models that guess instead of abstaining. Testing on SimpleQA-Verified shows that current volume-based accuracy metrics reward ungrounded generation.