summary: The authors consider the problem of learning a classifier in the setting where the classifier output influences the system in turn. They present an approach to trade-off classifier quality with “lookahead regularization”, i.e. potential beneficial impact of classifier prediction on system behavior. pros: - important aspect not often studied: impact of predictions on behavior - clearly written, instructive illustrations - relevant literature mostly discussed cons: - $\Gamma$ must be known: assumption that a causal model is available meta-review: To me, this is a borderline paper for the following reason: I am not convinced that the basic assumption (which underlies the whole paper), that the prediction method has to also serve to make intervention recommendations, is sensible. None of the introductory examples convinces me of this. Consider the wine example, where one task is to predict wine quality and the other is to improve wine quality with policy recommendations -- I do not see the necessity to conflate them, I think in practice it perfectly valid to output both: a quality prediction tool (potentially taking into account spurious correlations), as well as a causal analysis of what should be changed to improve quality. The latter is actually assumed known in the paper in the form of the masking $\Gamma$. In general, I think this paper could contribute in a non-constructive way to the already murky discussion on the relation of predictions and interventions. As all reviewers favourably assessed the paper positively and it is not technically flawed, it is accepted. However, I urge the reviewers to reflect and discuss in the paper the basic assumption that a single classifier should take on two different tasks of prediction and policy recommendation; just because this is current practice in some applied ML system does not mean that this is desirable. The paper could be made much stronger IMO if it could convince a reader like me, to whom this setting looks like “narrow framing” for the sake of technical tractability, without considering its justification nor plausible alternatives.