The teacher–student relationship is modeled as an agency problem, where teachers are concerned with human capital formation and students with ability signaling. We distinguish between two cases depending on whether in ability inference the job market can or cannot observe the grading rule applied. We show that many empirical grading patterns, including grade compression and inflation, are all consistent with optimal ability screening when grading rules are unobservable. With observable grading rules, the teacher perfectly screens students' abilities, provided that certain conditions hold. We apply the model to discuss policy applications such as “No Child Left Behind.”