The first criterion is that the model has to work. This is an engineering validation to see whether the theory is competent and complete enough to process the information required within the resourse and other constraints.
The second criterion is that the model should work across the full range of species, with gradual and step-wise additional functionality to handle the additional information processing requirements of later, more evolved species. Each of the step-wise enhancements should give an evolutionary advantage, and none should require a fundamental redesign.
The third criterion is that the model should utilize known processes as much as possible. We have listed randomization, selection, learning, adaptation, and mimicry as likely candidates.
Even meeting the above falsification criteria gives us little direct assurance that the information processing in the model is equivalent to the information processing in individuals of the species. However, being in a dynamic search space with ongoing optimization, if it can work, then it is likely that nature will have tried something like it.