Contrary to predictions of a stochastic self-terrninating search model, choice reaction times were lower for rulebound positive sets than for nonrulebound positive sets when memory set sizes were varied and stimulus probabilities were held constant. The result suggests that rule use and chunking, as opposed to probability alone, facilitate character classification. As a corollary, it is proposed that serial and parallel processing differ in degree rather than in kind. The extent to which the memory set has been chunked is the critical variable affecting Ss' scanning practices.In a character classifieation experiment, the S learns a list and then decides as quickly as possible whether single items presented later belong in the list. Generally , the time required to make the decision increases monotonically with the length of the list (Sternberg, 1966). This fact suggests that Ss compare the test item with each item in the memorized list in aserial fashion. One of the surprising discoveries made with the character classification procedure is that an equal amount of time is taken to indieate whether an item is a member of the list or not. Intuitively, one would expect judgments of membership to take much less time than judgments of nonmernbership, since the memory search would be expected to stop when a positive match is made between a test and list item. But because choice reaction times (CRTs) for judgments of membership and norimembership are roughly equal and can be represented by curves of equal slope, it seems that judgments are made after an exhaustive search through the memory set (Atkinson, Holmgren, & Juola, 1969;Briggs & Blaha, 1969;Sternberg, 1966).Objections have been raised against the claims that a serial and exhaustive search is used in character classification. The monotonie function predieted by a serial search model has not been obtained in several character classification experiments (Clifton, 1973;Kristofferson, 1972b). Sometimes the serial search model has been invoked only because predictions based on a parallel search model were too difficult to derive (e.g., Snodgrass, 1973). Many experiments using tasks other than character classification have yielded results which only a parallel search model could predict (Egeth, Jonidas, & Wall, 1972;Leonard, 1961; Morin, Koniek, Troxell, & McPherson, 1965;Neisser, 1963;Neisser, Novick, & Lazar, 1963; Travers, 1973 Tisseyre, 1966; Biederman & Zachary, 1970;Falmagne, 1965;Hawkins, Thomas, & Drury, 1970;Hyman, 1953;Orenstein, 1970) and sequential effects on reaction time (Bertelson, 1963(Bertelson, , 1965Falmagne, 1965; Leonard, Newman, & Carpenter, 1969;Remington, 1969;Schvaneveldt & Chase, 1969). Opponents of the exhaustive search model have proposed that scanning is carried out in a self-terminating manner: all items are stored with their respective positive or negative choiee responses; these paired associates are stochastically ordered in memory stacks according to their recency and frequency.In the present paper, the stochastic self-terminatin...