I used to 'sit in' on interviews, and basically say very little except to answer questions about working here or to serve as the 'linter' to get them unstuck.

You hear something very different than the person who asked the question. And in one case I discovered that my coworker had been asking a question that he thought he knew the answer to and did not; a deep clone implementation (where he never asked about cyclic graphs).

I assumed it rather than say it out loud, but my idea is to have standard, vetted questions, to reduce the risk of a question being mal-formed like that (unless you want to see how they make tradeoffs in response to an overambitious spec). Your situation sounds rather different. Do you think my scenario would work better?