Predicting the effect of unseen interventions is a fundamental research
question across the data sciences. It is well established that, in general,
such questions cannot be answered definitively from observational data, e.g.,
as a consequence of unobserved confounding. A generalization