The only criterion that makes sense to me is that a cochlear model should replicate well-known psychoacoustic experiments without having to apply "special adjustments" for each experiment. For example, the two-tone interference experiments such as missing fundamental, combination tones, masking, etc. should all be explainable by the same model. Likewise for many other phenomena, such as rippled noise, whispered speech, real-time pitch detection, or detection of tones without periodic repetition, not requiring exotic computations. Furthermore, the model should be capable of separating sounds in terms of awareness and attention.
As far as I know, no current models can come anywhere close to doing even one of these things. So you have an open field to pursue.
Here's a hint: Helmholtz was wrong, Seebeck was right.