The question of whether linguistic knowledge is binary (i.e., grammatical vs. ungrammatical) or stochastic is one of the most important questions in general linguistic inquiry. Much recent work in the last few decades has argued that phonological knowledge is stochastic (e.g., Hayes & Londe 2006). Building on this body of research, we show that in Japanese, gradient phonological knowledge affects several word formation patterns in stochastic ways. Concretely, we show that identity avoidance effects hold at both the segmental and the CV-moraic levels and stochastically affect two types of word formation patterns in Japanese: group name formation and rendaku. We show that Maximum Entropy Grammar (Goldwater & Johnson 2003), together with multiple OCP constraints (Coetzee & Pater 2008), successfully models both of the observed morphological word formation patterns without any further stipulation. In addition to this theoretical contribution, one of the patterns discussed in this paper—group name formation—has not been analyzed from the perspective of formal phonological theories before, and hence this paper has descriptive novelty as well.*