Tangles, as introduced by Robertson and Seymour, were designed as an indirect way of capturing clusters in graphs and matroids. They have since been shown to capture clusters in much broader discrete structures too. But not all tangles are induced by a set of points, let alone a cluster. We characterise those that are: the tangles that are induced by a subset of or function on the set of data points whose connectivity structure they are meant to capture. We offer two such characterisations. The first is in terms of how many small sides of a tangle's separations it takes to cover the ground set. The second uses a new notion of duality for oriented set separations that becomes possible if these are no longer required to be separations of graph or matroids.