In this paper we consider the problem of incorporating mechanisms for expressing abstractions and generalisations into constraint-based formalisms for linguistic description. Much recent work has been done on the denotational semantics of these languages, rather less on their suitability for the task of producing realistically large linguistic descriptions (their "phenomenological" semantics) and on the desiderata for their empirical computational behaviour (their operational semantics). In essence, an adequate formalism should have the properties both of a good programming language and of a support for theory construction.