The idea of pseudo-classification based on external relevance was introduced and compared with the usual classifications using associative techniques. A general model for an information retrieval system using term classification was described. A set of operators for adjusting pseudo-classifications and preventing their deterioration was derived for a particular match function conforming with this model. Pseudo-classifications were used for prediction of relevant documents and for the evaluation of retrieval systems with respect to their theoretical optimum. The concept of improvability of a retrieval model with respect to its constituent submodels was introduced and elaborated upon. (RR)