Abstract A nonfrequentistic model of subjective probability was tested in an empirical, subjective estimation task. The model relies on a characterization of stimulus similarities as distances in a cognitive space. This representation was independently tested using a multidimensional scaling analysis of similarities data. Subjective probability estimates given by the subjects for a related set of stimuli were then shown to be highly consistent with respect to their similarity judgments, and predictable by the model. The experiment also utilized special transformations on the stimuli to test consistency properties of subject probabilities.