Current thinking on inflection classes views them as organized networks rather than random assemblages of allomorphs (a view that reaches back to the 1980s, with such notions as Wurzel's paradigm structure conditions and Carstairs's paradigm economy). But we still find systems which appear to lack any visible implicative structure. A particularly striking example comes from Võro (a variety of South Estonian). Its system of verbal inflectional suffixes is formally simple but distributionally complex: although there are never more than three allomorphs in competition, nearly two dozen inflectional patterns emerge through rampant cross-classification of the allomorphs. Allomorph choice in one part of the paradigm thus fails to constrain allomorph choice in the rest, so it looks as if the paradigms would have to be memorized en masse. The key to these patterns lies outside the system of suffixation itself, in the more conventional formal complexity of stem alternations and their paradigmatic patterning. The computationally implemented analysis presented here provides a model of inflection in which the implicational network of phonological, morphophonological and morphological conditions on formal realization are unified in a single representation.