Automating Normal Science: Reusing Exemplars in Quantitative Experiments Gustavo Lacerda da Melo Abstract: We design and implement a system for making derivations in physics and engineering, by reusing knowledge from previous derivations. Many intermediate results found in physics and engineering derivations do not follow formally from the laws and antecedent conditions about the system, and according to a view called "particularism", this problem cannot be corrected. To deal with this, our reuse is divided into results that follow from the theory and antecedent conditions (neat reuse), and "tricks" that do not (scruffy reuse). An initial formalization is proposed and tested. An equational theoremprover is developed, and examples of reuse are run on it, following Bod's EBE model. The formalization is finally enriched to include semantic variable-naming, preconditions and axiom-tagging. It is proposed that this framework offers significant progress towards the goal of modeling scientific reasoning. We run the system with examples from the domain of classical mechanics. The reasoning system can be seen as either problem-solving AI or as a cognitive model of an idealized scientist, behaving according to our model of normal science. Philosophically, this corpus representation can be seen as a computational formalization of Kuhn's notion of exemplar, and the retrieval heuristics as a mechanism by which such exemplars get reused in normal science. This thesis is not available online. To obtain this thesis, please contact the author.