This paper seeks to explain the variation in the scale of violence across episodes of ethnic conflict, using data from Rwanda and Burundi. To do so, we explore the Òdark sideÓ of in-group policing - when it is exploited for genocidal killing, instead of being employed to moderate ethnic hostilities. Our efforts build upon Fearon & Laitin (1996), who concede this mechanism could backfire if an ethnic group announces its intent to attack, rather than cooperate with, a rival ethnic group. We depart from them in developing a computational model that assumes individuals vary in their propensity to engage in violence, form independent beliefs about nominal rivals, and respond to catalysts, namely messages about ethnic violence that has transpired. In addition, members of the politically dominant ethnic group are subject to metanorms, i.e., sanctions for non-participation. Given these reasonable assumptions, our model yields substantial variance in the scale of violence across episodes. We further demonstrate that (1) the interaction between nominal ethnic rivals is rarely deterministic and, consequently, an emphasis on structural factors is somewhat misguided; (2) changes in inter-ethnic trust influence the patterns of ethnic conflicts - communities exhibiting high levels of trust are more likely to experience intense violence that subsides rapidly, in contrast to the moderate violence that is sustained over a longer period of time in less trusting communities; (3) stronger metanorms engender more extensive episodes of violence.