What an agent knows: evaluation in open worlds