It will make output more consistent but not more accurate, the same wrong answer, more reliably.

It will reduce fabrication meaningfully, because lower randomness makes the model stay closer to its training data.

It will have no effect at all, since temperature only influences creative tasks rather than factual ones.

It will increase fabrication, because constraining word choice forces the model into overconfident phrasing.