Here is the failure this fixes, and it is the expensive one.
The model tells you something. It is specific, it is fluent, it has a number in it, and it is wrong. Not obviously wrong, which you would catch. Wrong in the way a confident colleague is wrong: a version number that is one release out, a feature the framework does not have, a dosage that sounds clinical, a case that was decided the other way.
So you do the natural thing and ask "are you sure?". Watch what that actually produces. Sometimes the model folds and apologises for a claim that was correct. Sometimes it restates the original with more conviction. Neither response is verification. Both are just continuations of a conversation, and the conversation currently contains a confident false sentence that the model is now conditioned on.
That is the mechanism worth understanding, because it also contains the fix.
A model can produce a plausible claim, and it can recognise whether a claim is true. It cannot do both in the same reasoning step.
Generating and checking are different jobs. When they happen in one pass, the generating wins, because the half-written sentence in front of the model is the strongest evidence available about what the next word should be.
Separating the two jobs
Chain of Verification, developed by Dhuliawala and colleagues at Meta AI Research in 2023, pulls them apart into four steps. Draft an answer. Turn every factual claim in that draft into a specific question. Answer those questions on their own, away from the draft. Then rewrite the answer so it agrees with what the checking found.
The gains are not cosmetic. On biography generation, a task that is almost entirely factual claims, FACTSCORE improves from 55.9 to 71.4. That is roughly a third of the remaining errors, removed by nothing but restructuring the request.
And there is one condition on all of it. The checking step has to happen independently, with the draft out of view. Let the model verify while reading its own draft and it copies the error forward, now with a tick next to it.