When analyzing a networked system from the outside as a "black box", it is generally a difficult problem to make any inferences about the internal structure of the network that the system consists of. I will present some general thoughts on how much one can expect to be able to infer, and outline a potential technique for making such inferences based on hypothetical rewirings of copies of subsystems. So far, the method has been worked out in detail only for the case of Bayesian networks, in https://arxiv.org/abs/1609.00672.