A parsimonious lexicon must encode generalisatious (e.g., ). One then needs to reason about when these apply. A general consensus is that an operation known as default inheritance is useful for this ([2, 4, 8, I0, II], and others). A frequent motivation for using it, is to capture the overriding of regularities by subreg'~t~xities in a computationally ei~cient m~nner. Information need only be stated once, instead of many times in each separate word, and default inheritance ensures that words inherit the right information. But there’s a problem with this. Many lexical generalisations are of the sortwhere there are exceptions to the rules, which are triggered by information which resides outside the lexicon. In particular, pragmatic knowledge can trigger exceptions and default inheritance doesn’t communicate properly with pragmatics to encode this. In this paper, we'll consider threexamples where this occurs: logical metonymy (e.g., enjoy the book means enjoy reading the book), adjectives (e.g., the interpretation of last in fast car, last motorT#ay, fast typist etc.), and noun-verb agreement. We'll argue for a new version of default inheritance, which allows default results of lexical generalisations to persist as default beyond the lexicon. We'll show that this persistence can be exploited by the pragmatic component, to reason about when generalisations encoded in the lexicon survive in a discourse context. We'll represent the link between the lexicon and pragmatics via two axioms. These will predict the pragmatic exceptions to lexical generalisations that arise in a discourse context. We thereby explain how words are interpreted in discourse, in a way that neither the lexicon nor pragmatics could achieve on their own.