Where this one gets misread
Retrieval quality and answer quality get reported as a single number. An answer can read well from the wrong passages, and a system can retrieve perfectly then summarise badly. Measured together, neither problem is visible.