As the number of computational models of eye-movement control in reading increases, so too will their coverage and complexity. This will make their comparison and testing increasingly challenging. We argue here that there is a need to develop a methodology for constructing and evaluating such models, and outline aspects of a possible methodology.