|Abstract:||This paper provides two arguments that constraint-based grammars should not be learned by directly mirroring the frequency of constraint violation and satisfaction in the target words of a language. The first argument comes from a class of stages attested in phonological development, called Intermediate Faith (IF) stages, in which children produce marked structures only in privileged positions. Two such stages are presented and analyzed, from the literature on English and French L1 acquisition, and their learning consequences are examined. The second argument concerns the degree of restrictiveness that a learner’s end-state grammar encodes, using two hypothetical interactions between learner’s assumptions about hidden structure and developing constraint rankings that can trick a learner into adopting a superset grammar. These two arguments are used to support an approach called Error-Selective Learning (ESL), in which errors are learned and stored gradually, in a way that relies on violation frequency, but rankings themselves are learned in a non-gradual way (relying on the algorithms of Prince and Tesar 2004; Hayes 2004). It is also shown that violation frequencies can still cause problems regardless of a learner’s method of grammatical evaluation – either ranked constraints as in Optimality Theory, or weighted constraints as in Harmonic Grammar.