For instance, if bees are getting afflicted by some fungus, then there is some research that says that applying fungicide X will allow the colony to prevent y% decline in the ideal case, or whatever.
At that point, even though the model will then incorporate this "sub-model," you don't need the system dynamics model to say what will happen when you apply the fungicide, you only need that well-researched sub-model.
I feel (and I'm getting a little out of my depth because of my limited research on real-world uses of these models) that the system dynamics model is only useful for attempting to determine what happens when you have a conjunction of various interventions, or interventions with various environmental parameters that were outside the original research of the sub-model. But that's precisely were the limitations of the system dynamics model will come into play, because the non-linear or chaotic effects may mean that the model's prediction that the colony will grow or fail may have little to do with reality, because of the complexity of accurately accounting for so many parameters, and guessing how they interact without specific research on how they interact.
I would describe all my work as "toy" models, in the sense that they are designed to demonstrate some aspect of an ecological system in a pedagogical way. The purpose of any one model is generally (1) to show 2 or 3 emergent phenomena, (2) to simply showcase the complexity of interconnected systems, where (say) eliminating a predator species can have surprising effects on the ecosystem at large, and (3) to introduce students to the idea of modeling as a practice -- i.e. making it explicit that they are working with a model, that models are not reality but make certain assumptions and decisions, and that the students themselves may be involved in making those decisions, and needing to defend them (from some work with Uri Wilensky, I also know that this is explicitly why NetLogo models always allow the end-user to edit the code, and this cannot be turned off).
I would describe a credible ecosystem model as one from which justifiable predictions can be drawn. These would need to be demonstrated experimentally.
Personally I don't know any credible ecosystem models, because that's not my field, but I'm sure that some exist, to varying degrees of credibility. My skepticism of the Bee model was based more on my knowledge of how high that cliff would be to scale, and how this model seemed not significantly more complex than plenty of models I have written.