A new approach to the student modeling problem is based on the idea of representing subject matter knowledge in sets of constraints. Constraint violations on the part of the student indicate incomplete or incorrect knowledge and can therefore be used to guide the response of an intelligent tutoring system. This approach promises to eliminate the need for runnable models of either the expert or the student and to reduce the computations required for student modeling to pattern matching. An application in the domain of subtraction illustrates the feasibility of the concept.