This article provides experimental evidence for the claim in Hayes (2004) and McCarthy (1998) that language learners are biased to assume that morphological paradigms should be phonologically-uniform--that is, that derived words should retain all the phonological properties of their bases. The evidence comes from an artificial language word-learning paradigm, in which children learned novel objects names such as "wutch" and a plural suffix -"del" in an alien language and then were asked to produce alien words with difficult coda-onset clusters, some of which straddled the singular +"del" suffix boundary. The results suggest that 4-year-olds who are acquiring novel consonant clusters are preferentially faithful to the base segments in a derived word, e.g. "wutch" in plural "wutchdel." The article interprets these results from the perspective of Optimality Theory (Prince & Smolensky 2004; Tesar & Smolensky 2000), using Output-Output Faithfulness constraints (Benua 2000) to understand the asymmetries in the observed phonologies of derived and underived test items. (Contains 2 figures, 23 tables, and 31 footnotes.)