The models are constrained by what is known at the time they are built - those that clash with existing observations are discarded while those that don't, get published.