A language learner lacking systematic negative evidence could rely on the Subset Principle (SP) to avoid adopting an overgenerating grammar. But SP will be effective in fending off overgeneration only if learners are able to determine, reliably and across the board, whether two candidate grammars stand in a subset/superset relation, and if so, which of them gives the subset language. Whether this is easy or difficult depends in part on the assumptions we make about the form in which grammars are mentally represented. I take it as obvious that learners do not assess potential subset relations by generating languages and comparing them; rather, they must be able to settle the matter by comparing grammars. 1 So learning would be facilitated if subset relations between languages were reliably associated with some readily accessible relationship between grammars. On traditional hypothesis-testing models of learning, which assumed rich grammars needing to be constructed from scratch, no such transparent relationship could be identified; that is, the search for an evaluation metric was a failure. The parameter setting model offers a 430welcome solution. Because parameters and their values are assumed all to be innately specified, it can be assumed also that whenever two values of a parameter stand in a subset/superset relation, they are innately ordered, with priority given to the subset value. Then a learner can satisfy SP very easily, just by obeying the injunction: Select the earliest (highest priority) value of a parameter that is consistent with the available evidence.