The procedures of classification employed in the observational sciences, particularly in psychiatry, are situated with respect to two others: those which are utilized by the cognitive instrument and those of the logician. The problems posed in the observational sciences arise from the necessity of rendering explicit procedures which are used in situations inevitably presenting a certain degree of incertitude. A stochastic model of the patient-as-seen-by-the-observer is proposed. It is emphasized the prognostic assessment may be carried out without relying on any classification whatever, by means of an equation of multiple regression making use of symptoms as predictors. One then offers to explain why psychiatry strives to categorize symptoms and patients. Two problems are examined in the light of this discussion. The first one is whether it is better to work out a unique system of classification or to favor a plurality of such systems. In light of the preceding discussion, the latter alternative is to be preferred, whereby the notion of 'natural classification' is rejected. The second one has to do with the factors creating incertitude in the observer using a given system of classification.