We are implementing an automated theorem proving system that will in part attempt to simulate human per-formance on calculational logic. To support this project, we recorded and analyzed people's eye movements while they constructed calculational proofs. Our find-ings confirm some expected behaviors (based on strate-gies and principles taught to students) that previously may have seemed untestable, such as the influence of the form of the current proof step and of syntax in gen-eral on the moment-by-moment computations of the problem solver. The experiment also uncovered other interesting patterns, such as the seemingly inefficient but widely occurring tendency to attend to particular premises despite their not being used in the proof under consideration. Overall, we gained insights into real-time problem solving that directly apply to our automated system and could not have been gained merely by studying written proofs. Our findings also demonstrate that analyses of eye movements can improve our under-standing of the psychology of theorem proving.