Yeah, that's the obvious counterexample, but that's why I said "more often than not". The statement I was questioning was that fast-math is "normally not more correct."