If I remember it correctly, later behaviorists said that it is also scientific to propose the existence of "hidden variables", if the model needs it -- as long as one does not suggest what exactly they are, because that would be unscientific.

These behaviorists could explain an angry person by saying that punching them in the nose activated a hidden variable X, and that the hidden variable X can make the person punch back, or kick, or do something like this. It is all correct, as long as you avoid connecting the hidden variable X with anger (because anger is unscientific).

This kind of model could get just as close to reality as any other model, only without the obvious names for concepts. Which perhaps could be useful for building an artificial intelligence, because we would not make the mistake of believing that just because the variable has the correct name, it must work correctly.