Summary: The conventional preference disaggregation approaches for multiple criteria sorting aim at reconstructing an entire set of assignment examples provided by a Decision Maker (DM) with a single preference model instance. In case the DM’s holistic preference information is not consistent with an assumed model, one needs to accept that some assignment examples are not reproduced. We propose a new approach for handling inconsistency in the context of a threshold-based value-driven sorting procedure. Specifically, we introduce preference disaggregation methods for reconstructing all assignment examples with a set of complementary preference models. The proposed approach builds on the assumption that the importance of particular criteria or, more generally, the shape of marginal value functions and their maximal shares in the comprehensive value are contingent (i.e., dependent) on the performance profile of a given alternative. Therefore, in case of inconsistency, the set of assignment examples is divided into subsets, each of which is reconstructed by a unique model to be used only if certain circumstances are valid. We present three methods for learning a set of contingent models, allowing different degrees of variation in the contingent models along two dimensions: the shape of marginal value functions and interrelations between the models. To apply such a set for classification of non-reference alternatives, we learn a decision tree which makes the application of a given model dependent on the alternatives’ profiles represented by the performances on particular criteria, hence allowing to select an appropriate model among the competing models to evaluate a non-reference alternative. The method’s applicability is demonstrated on a problem of evaluating research units representing different fields of science.