We studied the influence of piscivorous fishes and prey refuges on assemblages of fishes occupying 52 model reefs in a large seagrass bed off St. Thomas, U.S. Virgin Islands. We conducted three experiments: two involving 6 reefs each, lasting 2 and 5 yr, and one involving 40 reefs, lasting 1 yr. Each experiment included replicate reefs in various combinations of five structural treatments: holeless controls, 12 and 24 small holes, and 12 and 24 large holes. Tagging studies indicated that the reefs were sufficiently isolated from each other to comprise statistically independent replicates, and that resident piscivores occupied home reefs. We observed 97 species on or near the reefs, representing all major foraging guilds, and each holed reef supported hundreds of individuals. We examined four categories of fish: (1) large reef associates (too large for the small holes; most of these fish were both predators on smaller fish and prey for larger transient piscivores), (2) moray eels (piscivores that could fit into the small holes), (3) small reef associates (potential prey that could fit into the small holes), and (4) juvenile grunts (potential prey that sporadically were extremely abundant). We tested five a priori predictions of the general hypothesis that predation is an important process structuring reef-fish assemblages. The first two predictions dealt with the role of prey refuges. First, if reef holes function as prey refuges, then prey fish should be most abundant on reefs providing holes near their body diameters, because such holes would make the prey fish safest from predation. Seven of eight experimental comparisons supported this prediction, and five of them were statistically significant. Second, if refuge availability limits prey abundance, then prey fish should be more abundant on reefs with 12 holes than those with no holes, and should be more abundant on reefs with 24 holes than those with 12 holes. The first part of this prediction was verified by all nine experimental comparisons, seven of which were statistically significant. However, there were no strong differences between 12-hole and 24-hole reefs. Thus, between 0 and 12 holes per reef, holes limited local prey populations; between 12 and 24 holes per reef, the number of holes was not limiting. Several lines of evidence suggested that the latter pattern was due to temporary saturation of the study area with refuges when we added 40 reefs to 12 existing reefs. The remaining three predictions dealt directly with the community-level role of predation. First, predators should affect local prey abundance either chronically, in which case a negative relationship among reefs is predicted between the average abundances of predators and prey, or sporadically, in which case a negative relationship is predicted between the abundance of predators and the maximum number of co-occurring prey ever observed at each predator abundance. The former prediction was falsified, whereas the latter was verified. Observations of extreme type III survivorship of recruit cohorts on reefs with many piscivores and occasional direct observations of piscivory bolstered the conclusion that this relationship was causal. Finally, we predicted that predators should affect the number of prey species on a reef. We observed a significant negative relationship among reefs between predator abundance and maximum prey-species richness. Comparing species' relative abundances on reefs at the extremes of this regression, piscivores appear to have nonselectively reduced and extirpated both common and rare prey species, although this relationship remains purely correlative. In our model system, high local species diversity appears to have been maintained despite rather than because of predation. We propose a conceptual model where the local abundances of coral-reef fishes are determined by the relative magnitudes of recruitment by larvae, colonization by juveniles and adults, predation, and competition for refuges, each of which varies through time and space. Multifactorial field experiments will be necessary to test such pluralistic hypotheses.