In California it's illegal to have a pet gerbil; in other states it isn't. You wouldn't for one moment cite this as an argument that California is a country. Or would you?

Whether we ultimately consider California a country or not is just an argument about the meaning of words (and the practical answer is that we have the word "state", which usually means a country but also means California). But I'd certainly say that the US is a very noncentral example of a country, and I'd warn people travelling there that the states of the US have some of the properties of countries and therefore it's important to e.g. check state laws in a way that you wouldn't do for subdivisions of more typical countries.

On the other hand, here is an empirical question where the models do in fact differ: Alice's model predicts that Dubai is a member of the UN and that the UAE (being nonexistent) isn't, while Bob's model predicts that the UAE is a member and that Dubai (being part of the UAE) isn't. Which model's prediction is more accurate? Or how about this: which entity has embassies in other countries? Also an empirical fact. Which model predicts it correctly?

Well of course the model that contains the correct the international-law technicalities is the model you'd want to use if you wanted to predict international-law technicalities. Just like if you want to predict where to find a tomato in a biology textbook, you should model it as a fruit. But if you want to know what to cook with it, you're better off modelling it as a vegetable.