A precise formulation of the notion of a rule in terms of sets and functions is proposed. It is argued that this molar formulation cannot be captured by networks of associations unless one allows associations to act on (other) associations. This formulation is then used as a basis for showing how rules are involved in decoding and encoding, symbol and icon reference, and higher order relationships. Decoding and encoding are shown to involve insertion into and extraction from classes, respectively. Reference is viewed in terms of rules which map equivalence classes of signs into the classes of entities denoted by these signs. Symbols are shown to involve arbitrary reference, whereas icons retain properties in common with the entities they denote. Higher order relationships are then expressed as higher order rules on rules. This is a direct generalization of associations on associations. Finally, a partial solution is posed to the vexing problem of "what (rule) is learned." Given a rule-governed class of behaviors, "what is learned" is defined as the class of rules which provides an accurate account of test data. Empirical evidence is presented for a simple performance hypothesis based on this definition.