Part of the Craig data have been set aside for evaluation. These are the test data. A disinterested third party will choose examples from the test data that can be used to appropriately test my hypotheses.
In particular I predict that Covlan will account for all the visual elements determined in the test data, where the elements found in the test data will be found using the same reliability methods used to get elements from the training data. This may seem at first a claim more about the consistency of the data than the coverage of my theory, but recall that analysis of the training data is not the only constraint going into the creation of Covlan. Realistically, I don't expect it will account for all of them, but the success of the theory on this measure is a function of its coverage.
If the primitives in the theory are used to represent multiple problems such that problem solving can occur, this will support the hypothesis that the primitives chosen are not case specific, and provide a general language for problem representation. I have four examples to model: The fortress/tumor problem, the Maxwell case, the furnace/factory problem, and the lab/weed-trimmer problem (the last two being from the Craig data). These four cases are very different from each other in terms of content, and constitute a representative sample of the physical systems domain that I am making claims about. Thus, my results should generalize to other physical problems.