By training it on a whole bunch of examples with valid external citations, to the point where it's able to hallucinate something that's valid. Of course, it won't always work, which is the point of TFA.

I don't think this is an answer. I'm basically claiming there is an uncontrollable error probability, and I think you agree with that. The person I'm replying to implies it can always be reduced (maybe even to zero).

Ah, yes, I see. The distinction between a working clock and one that's right twice a day.

I assume by forcing them to validate their citations with a deterministic tool.

What deterministic tool will validate that a citation faithfully represents the cited claim?