Also on the edge, but it appears they are relying on the search-augmented identification of conflicts in the generated statement, which is an easier task than constructing an answer to the question. It also encourages abstention because there are no conflicts in “I don’t know” (so “mitigating hallucinations” and “answering more questions correctly” are not necessarily the same thing)