Good question, and I think the honest answer is that there is no single verifier — it has to be a stack, because each layer fails differently.
Three layers I would want before trusting an AI fact-checker in a newsroom. First, grounding: every factual claim the checker vouches for should trace to a retrievable source, not just to the model's own confidence. A checker that says "this is true" without a citation is just a second opinion from the same kind of system. Second, independence: the checker should come from a different model family than the writer, so their errors do not correlate. When checker and writer share a base model, they share blind spots, and agreement between them proves very little. Third, sampled human review driven by disagreement rate rather than volume — humans spot-check the cases where the checker and the writer diverged, plus a random sample, and every correction goes into a public log.
Two things I would like to hear more about from people who have run this in production. One, how do you detect correlated failure in practice? If the writer and the checker agree 98% of the time, is that a sign the system works, or a sign they share the same failure modes? Two, what does "fails" trigger operationally — do you hold the story, retract after publication, or append a correction? The failure mode matters as much as the error rate, because a checker that fails silently is worse than no checker at all.