Within the framework of optimality theory (Prince & Smolensky 2004), Steriade (2001a,b) proposes the P-map hypothesis, whose fundamental tenet is that the rankings of faithfulness constraints are grounded in perceptual-similarity rankings. This article provides empirical support for this hypothesis. In Japanese loanword phonology, a voiced geminate, but not a singleton, devoices to dissimilate from another voiced obstruent within a single stem. Based on this observation, I argue that the [ voice] feature is protected by two different faithfulness constraints,IDENT( voi)Sing and IDENT( voi)Gem, and they are ranked as IDENT( voi)Sing » IDENT( voi)Gem in Japanese. I further argue that this ranking is grounded in the relative perceptibility of [ voice] in singletons and geminates, and this claim is experimentally supported. The general theoretical implication is that phonetic perceptibility can directly influence patterns in a phonological grammar.