Models could be roughly categorized as analogous, semantic, and abstract.
General models should be frequencionist and NOT probabilistic. The notion of probability is only applicable when we know exactly all possible outcomes or causes, which is, aside from a simplest systems such as a coin or a dice, is never obtainable or even fully observable. Probability is just a what proportion of this particular outcome takes on average? For most settings we cannot know.
To be precise - probabilities are applicable only to systems with could be defined as closed over a finite and fully observable set of possible states (outcomes). Whatever works for a deterministic discrete signal processing does not work for a stochastic, partially observable, continuous environments of physical reality.
Estimated probabilities bullshit, together with fancy topological bullshit is what ruined today's modeling. Nature is much simpler that our abstract, based on naive extrapolations models of it.