Interesting. The OP got me thinking along the lines of manually tagging salient features of each model (as well as ranking models by salience, either manually or automatically based on criteria related to the object that the model represents).