I really like the ending thought on "good enough". Trying to develop a model and verify it, the "good enough" point is crucial to know when to stop. As a physicist, I would add one more thing. It is really good if a model can contain at least some "physics" in addition to statistics. In the reading case that would be tying the model to theories of cognition. I made many models where the physics provided some basic constraints on the functional form, while statistics on actual data filled in the detailed calibration. I really preferred hybrid models of that sort. Especially when the model is used for the dangerous task of extrapolation.
I really like the ending thought on "good enough". Trying to develop a model and verify it, the "good enough" point is crucial to know when to stop. As a physicist, I would add one more thing. It is really good if a model can contain at least some "physics" in addition to statistics. In the reading case that would be tying the model to theories of cognition. I made many models where the physics provided some basic constraints on the functional form, while statistics on actual data filled in the detailed calibration. I really preferred hybrid models of that sort. Especially when the model is used for the dangerous task of extrapolation.