Behind a recommender system is the policy of the platform that runs it, which specifies acceptable system output and behavior. Policy is the language in which we communicate in order to reach consensus on what constitutes a harm-aware recommender system, and the measuring stick that allows us to enforce that consensus completely and consistently. Recently, policy has come to the forefront, as mainstream newspapers report that companies with large online platforms are limiting harmful items (YouTube) or removing items completely from their collection (Amazon). This talk discusses ways in which recommender systems can use algorithms to more closely connect with policies, allowing for better oversight and enabling harm-aware recommender systems. Our main example is a case study from bol.com, the largest e-commerce company in the Netherlands. At bol.com, policy enforcement is the responsibility of a quality team, who monitor the items that are ofered by third party vendors on the platform. We discuss the promise and problems of data programming, a hybrid human-AI paradigm, to enforce policy at large scale and change its enforcement quickly in response to policy updates. This talk aims to provide an interesting perspective on the relevance and potential of policy-related recommender system research.