In recent years, arguments that optimization within OT must proceed in two directions (production-directed and comprehension directed) have come from various sources. These include asymmetries in the linguistic abilities of children, learning algorithms, lexical blocking effects, freezing effects, superiority effects, and anaphora. These phenomena tend to be 'soft' in the sense that preferences can be overridden by context. I consider here the distribution of active and passive in a number of languages which are characterized by relatively little surface configurationality and argue that a satisfactory account requires bidirectional optimization. In such languages, diathesis displays a highly structured pattern of interpretational preferences and blocking effects, a pattern which can be characterized if optimization is bidirectional, but not if it is unidirectional. Interestingly, these effects are generally resistant to context and plausibility. Voice thus provides evidence of a clearly syntactic sort for bidirectional OT.