A chord-priming paradigm was used to test predictions of a neural net model (MUSACT). The model makes a nonintuitive prediction: Following a prime chord, expectations for the target chord are based on psychoacoustic similarity al short stimulus onset asynchronies (SOAs) but on implicit knowledge of conventional relationships at longer SOAs. In a critical rest, 2 targets were selected for each prime. One was more psychoacoustically similar to the prime, and the other was more closely related on the basis of convention. With an SOP, of 50 ms, priming favored the psychoacoustically similar target; with SOAs of 500 ms and longer, the effect reversed, and priming favored conventional relatedness. The results underscore the limitations of models or harmony based on psychoacoustic factors alone. These studies demonstrate how neural net learning models that are appropriately constrained can be subject to strong empirical verification.