Existingautomatedmodellingsystemseitherrely on large, complexlibrariesor requirecompleteaccessto the modelledsystem'sbehaviour,neitherof whichis desirable,To addresstheseproblems,a simplerarchitecturefor modelling knowledgeis described,basedon the separationbetween ideal modelsof componentsand correctionsthat can be applied to theseidea1models. The use of this architectureto develop accuratemodel boundariesis described,basedon considerationof interactionswithin such ideal models. A novel algorithm for refining models is also proposed. This algorithm considersbehavioural differencesbetweenmodels and appliesthe correctionsthat causethe greatestdifferencesin behaviour. Finally, somemodelsgeneratedby this methodareshownto be parsimonious.