Taxonomy-Driven Failure Analysis for Medical RAG Factuality Verification

This framework decomposes medical hallucination detection failures into retrieval-stage errors across five dimensions and verifier-reasoning errors across six stages. It addresses the lack of gold-standard evidence in open-ended clinical evaluation by providing structured diagnostics for retrieval-augmented generation.