282Working within a framework of exemplar-similarity models of memory (see, e.g., Estes, 1994;Kahana & Sekuler, 2002;Kahana, Zhou, Geller, & Sekuler, 2007;Medin & Schaffer, 1978;Nosofsky, 1986;Nosofsky & Kantner, 2006), we used sinusoidal luminance gratings as stimuli in a modified Sternberg (1966) recognition task. The metric properties of the grating stimuli were exploited to test a novel prediction generated by combining the exemplar-similarity approach with an explicit, signal detection account of decision making (Wickens, 2002).In exemplar-similarity models of recognition memory, it is assumed that a summed-similarity computation is a basic component of subjects' recognition judgment. This computation sums-over all study items-the p's similarity to each of the study items. According to the model, when this sum reaches or exceeds some critical value, the subject will say "yes," judging that the p had been among the n study items that had just been seen. Following convention, we will use the term target (T ) to designate trials on which p replicated a study item and the term lure (L) to designate trials on which p did not replicate any of the study items. On average, the value of summed similarity on T trials will exceed that on L trials, which means that P(yes) responses on T trials will be higher than those on L trials. The nature of the elements entering into the computation will also tend to produce a systematic difference in the variances of summed-similarity values on T and on L trials, which leads to an unexpected prediction for the slope of z-transformed receiver operating characteristics(zROCs).On T trials, values of summed similarity arise from two quantitatively different sources that differ in their respective variability. The first, far larger source of variability reflects the contribution of the n 1 study items that are not replicated by p. Random selection of study items from a stimulus pool means that some of n 1 study items will be similar to p, and that others will be very different from p. As a result of this random divergence, these n 1 nonmatching study items will contribute a highly variable amount of similarity to the summed-similarity signal for any trial. The second, smaller source of variability in summed similarity on T trials reflects the contribution of the one study item that the p does replicate. Over trials, this study item's representation will tend to be perceptually similar to p-even with the memorial noise postulated by the model. Because that study item and p are physically identical, they are likely to be perceptually similar, despite the random noise associated with the study item's memorial representation. As a result, similarity between this study item and p will vary over a narrow range clustered near 1.0 (Zhou, Kahana, & Sekuler, 2004). In three experiments, we examined connections between item-recognition memory and memory for itemposition information. With sequences of compound gratings as study and probe items, subjects made either itemposition judgments (Experim...