training on verifiable research seems like the right constraint. what kind of claim still trips it up most?