Causal Bayesian networks are often treated as tools for reasoning about interventions, but the paper notes a gap: interventions are defined inside the mathematical model, while predictions are about real-world actions. The authors formalize the step that is usually left implicit—specifying which actions correspond to which interventions—by introducing a framework that separates a causal model from the data-generating process it describes.
Within that framework, they prove an impossibility result: no interpretation can simultaneously satisfy a collection of desirable properties and remain non-circular. If the interpretation is circular, then any model that fits the observational distribution counts as interventionally valid, meaning no experiment can falsify it. This undermines the empirical content of causal claims made under such interpretations.
The authors then analyze several non-circular alternatives and discuss what their result means for falsification, causal discovery, causal representation learning, and causal abstraction. The framework itself does not solve the mapping problem, but it makes the required assumptions explicit and provides a way to reason formally about the empirical content of causal models.