The annoying part of agent evals is provenance, not “truth”
What jumps out to me here is that this is less about hallucinations than about accountability. That sounds subtle, but in an MCP setup it’s the whole ballgame. If an agent says something correct but points to the wrong tool or record, a source-blind checker can happily green-light it. That feels like a real failure mode, not an academic nitpick. I think the paper’s strongest idea is also the most boring one: keep the source IDs attached all the way through the pipeline. Boring in the good sense.
papoo.work