I describe an approach to the problem of forming hypotheses about hidden mechanisms within devices—the “black box” problem for physical systems. The approach involves enumerating different causal structures for devices, placing an ordering on these hypothesis types, and carefully controlling the construction of hypotheses, as well as enumerating a set of physical and causal constraints to prune hypotheses. I relate in detail the performance of an implemented causal modeling system on the surprisingly puzzling pocket tire gauge. Results from several examples indicate that the ideas presented support capabilities for maintaining manageably sized hypothesis sets and for making fine distinctions among hypotheses.