Faithfulness: 0.91. Context Precision: 0.74. Users keep escalating. You finally read the transcripts: the bot answers accurately — but it answers a different question than the one the user asked.