[gmueckl]

It goes deeper than that. If you can tweak your new model to correspond to an older one for which you have results available, you must run tests in which you apply these tweaks and verify that you get the same results. If you get different results, one of the models simply isn't doing what you think it's doing and you need to understand that. And in the vast majority of cases, it's likely a mistake in a model, not any kind of new science.

One pretty general debugging/analysis strategy for a complex model is to deactivate it piece by piece until you get it to a point where it should reproduce a result that you can verify by other, independent means. E.g. tweak the parameters of the model so that reproduces a result that is known analytically. For instance, whenever it is possible in a model, a good sanity check is to set parameters that must lead to e.g. a perfect conservation of certain quantities.