Rather, we concluded that we should hash input and output products, and write down that hash when we compute products. That way, when we go to do the initial schedule, we can do a very fast traversal of the graph out to the frontier just by hitting the artifact cache and asking it only for the output hash, not the full product