Rethinking Realist Evaluation of Complex Social Interventions: From Causal Snapshots to Adaptive Causal Learning

Realist evaluation using Context-Mechanism-Outcome (CMO) configurations offers a compelling account of generative causation in complex interventions. However, the theories of change within which CMO configurations are embedded often function as epistemic artifacts that fix causal claims and constrain adaptive learning. This gap is structurally reinforced by commissioning frameworks, accountability requirements, and reporting structures that encourage static representations of change. Drawing on a program-level theory of change developed for a multi-site randomized evaluation of community-based violence intervention across five UK cities, we show how complexity-aware realist evaluation can address these shortcomings. Tracing targeted police enforcement across delivery sites, we identify three dynamics—contextual recursion, temporal non-linearity, and configurational causation—that required progressive revision of the causal model. We distill this into six transferable principles that retain realist generative explanation while treating causal understanding as provisional and subject to iterative refinement. Although developed within realist evaluation, these principles apply more broadly to theory-based evaluation.