A family of previous findings suggests the working of a learning bias here dubbed domain generalization bias, or learner tendency to favor phonological constraints that hold across morphological domains (Martin 2011; Myers & Padgett 2014; Chong 2016, 2017). This paper argues, despite these findings, that domain generalization bias can be defied, and that the matching of phonological drives across morphological domains is not a necessary condition for learning. Evidence for the claim comes from a corpus study of Malagasy, which displays contradictory markedness preferences across domains: backness dissimilation applies regularly to the passive imperative suffix, but stems in the lexicon show an overall opposing preference for backness harmony. Though learners might prefer for morphological domains to match, the Malagasy system suggests they are capable of overriding this bias when they have access to natural constraints with opposing drives. Contradictory markedness preferences are even found to regulate morphemes in the same domain, in Yucatec Maya: as discussed previously in Krämer (2001), two suffixes in the language harmonize for backness and height, but one suffix dissimilates for backness, and yet another dissimilates for backness and height.