This article draws on two recent traditions in evaluation methodology, one grounded in complexity theory and the other in a realist philosophy of science. Sometimes seen as incompatible, it is argued here that complexity theory and realist evaluation are 'natural bedfellows'. Because of their similarities and differences, evaluators can usefully draw on both perspectives within the same evaluation. One way to do so is to draw on key concepts from each in the selection and use of substantive theory. The article introduces the concept of 'complexity-consistent' substantive theories and suggests that this is useful for the evaluation of policies and programs in complex adaptive systems. It demonstrates how substantive theories can be analysed in complexity terms, how multiple theories can be 'layered' to reflect multiple levels of systems, and how such theories can be used within evaluation design and analysis.