|Abstract:||Learners face potential superset traps when presented with a target language which allows some marked phonological structure only when that structure occurs as the outcome of a morphological mutation process. For example, the vowel-tensing ablaut which marks elatives in Javanese can create tense vowels in closed syllables, even though such vowels are otherwise not allowed in the language. In this paper, I show that restrictive final grammars will be learned in these situations, under the assumptions that mutation processes result from special faithfulness to floating-feature affixes (Zoll 1996) and that learners have a Specific-F >> General-F bias (Smith 2000). I also show that such languages create a paradox for a purely declarative account of affixation: end-state restrictiveness in Javanese requires morpheme-realization constraints to be biased towards a ranking above IO-Faith, in tension with arguments by Adam & Bat-El (to appear) that data from child Hebrew require the opposite bias. I also show that the tension between the Javanese and Hebrew facts disappears under an item-based theory of morphology, where all affixation (including of the mutation variety) results from faithfulness to underlying structures.