|Abstract:||Many constraint-based accounts of language-internal phonological inconsistencies appeal to the idea that the phonological lexicon is stratified into distinct subcomponents. This article adopts the perspective of the language learner in reexamining two assumptions related to this approach: 1) that the membership of any lexical item can be determined on the basis of distributional evidence (strong lexical stratification); and 2) that ranking inconsistencies should be modeled through multiple faithfulness constraints indexed to different strata (indexed faithfulness). Using the case of post-nasal voicing in Japanese, it is argued that a phonological grammar with strong lexical stratification cannot be acquired given the lack of positive distributional evidence available to the learner. When the strong lexical stratification hypothesis is retracted, a computational problem emerges for the acquisition of indexed faithfulness. Taking these issues into consideration, the article explores a revised model of lexical stratification that can be learned based on alternation data and general learning mechanisms.