This article examines two models of scientific reasoning (dogmatic and sophisticated falsification) as metaphors for the problem-solving activities of a specific work role. Think-aloud protocol data were used to identify problem-solving methods of two electronics experts performing process diagnosis in a very large scale integration (VLSI) fabrication environment. The expert with more experience used sophisticated falsification whereas the expert with less experience used both dogmatic falsification and sophisticated falsification. Computational models (production systems) were constructed and compared with performance of the experts on a sample of diagnostic problems. A neural network was also constructed to examine the development of the two reasoning models. The developmental study suggests that sophisticated falsifications may be a "naturally" occurring metaphor for diagnosis under the assumption of expectation-based learning and environments of complexity and change. (Imran Zualkernan and Paul Johnson)