We present a novel protocol, M3L, for multicast tree fault isolation based purely upon end-to-end information. Here, a fault is a link with a loss rate exceeding a specified threshold. Each receiver collects a trace of end-to-end loss measurements between the sender and itself. Correlations of loss events across receivers provide the basis for participants to infer both multicast tree topology and loss rates along links within the tree. Not all receiver traces are needed to infer the links within the network that exceed the loss threshold. M3L targets the minimal set of receiver traces needed to identify those loss-exceeding links. While multicast inference of network characteristics (MINC) is well understood, the novelty of M3L lies in the manner in which only a subset of the receiver traces is used. We model this as a problem of establishing agreement among distributed agents, each acting upon incomplete and imperfect information. Considering bandwidth to be a limited resource, we define an optimal agreement to be based upon the smallest possible receiver set. M3L performs well compared to the optimal.