Why similar LLM agent scores need different fixes

A practical document evaluation that separates missing evidence from wrong answers, with AgentHop as the starting point.