Abstract The discovery and validation of knowledge representations for new types of reasoning is a vital step in artificial intelligence (AI) research. A clear example of this process arises in our recent study of expert physicians' knowledge of the physiological mechanisms of the body. First, we observed that the reliance on weighted associations between findings and hypotheses in first-generation medical expert systems made it impossible for them to express knowledge of disease mechanisms. Second, to obtain empirical constraints on the nature of this knowledge of mechanism in human experts, we collected and analysed verbatim transcripts of expert physicians solving selected clinical problems. This analysis led us to the key aspects of a qualitative representation for the structure and behavior of mechanisms. The third step required a computational study of the problem of inferring behavior from structure, and resulted in a completely specified and implemented knowledge representation and a qualitative simulation algorithm (QSIM). Within this representation, we built a structural description for the mechanism studied in the transcripts, and the simulation produced the same qualitative prediction made by the physicians. Finally, the system is validated in two ways. A mathematical analysis demonstrates the power and limitations of the representation and algorithm as a qualitative abstraction of differential equations. The medical content of the knowledge base is evaluated and refined using the standard knowledge-engineering methodology. We believe that this combination of cognitive, computational, mathematical, and domain knowledge constraints provides a useful paradigm for the development of new knowledge representations in artificial intelligence.