|Abstract:||It has been observed that grammaticality judgments do not necessarily reflect relative corpus frequencies: it is possible that structure A is judged as more grammatical than structure B, whereas at the same time structure B occurs more often in actual language data than structure A. In recent work (Boersma & Hayes 2001), we have used Stochastic Optimality Theory to model grammaticality judgments in exactly the same way as corpus frequencies, namely as the result of noisy evaluation of constraints ranked along a continuous scale. At first sight, therefore, this model seems not to be able to handle the observed facts: linguistic forms that have zero corpus frequency due to harmonic bounding often turn out not to be totally ungrammatical (Keller & Asudeh 2002), and 'ideal' forms found in experiments on prototypicality judgments often turn out to be peripheral within the corpus distribution of their grammatical category (Johnson, Flemming & Wright 1993). In this paper, I argue that the paradox is solved by assuming a listener-oriented grammar model (Boersma 1998), in phonology as well as in syntax. In that grammar model, the natural way to derive (relative) corpus frequency is to measure the production process, whereas grammaticality judgments naturally derive from a simpler process, namely the inverted interpretation process.