To gauge the success of a model-theoretic semantics, we generally look first to the set of logical truths it produces, to the sentences that turn up true in all models. If we've got things right, these sentences ought to conform fairly well to our pretheoretic intuitions about the logical or analytic truths of the language in question. Or, what comes to much the same thing, we look to the relation of logical consequence that emerges from our semantics. When the semantics declares one sentence to be logically consequent upon some others, the corresponding inference should seem - again bringing to bear our pretheoretic intuitions - to be a valid one. In this paper, I discuss the import of this criterion on model-theoretic semantics.