Rimac is a natural-language tutoring system that engages students in dialogues about physics concepts. It implements empirically-derived decision rules to guide the tutor’s questions and responses to student input. These rules simulate the highly interactive nature of human tutoring. Our aims are to identify particular dialogue patterns whose frequency predicts learning and determine if these relationships vary by student characteristics (e.g., prior knowledge, gender). We will demonstrate the rules “in action” and summarize the results of a study in high school physics classrooms that compared an experimental version of Rimac that deliberately executes these rules with one that does not.