Before the fact or after: when a safety property gets enforced
Two fixes can share every technical detail and still differ in the one thing that matters most: whether the safety property gets enforced before a policy is chosen, or repaired after the fact. Concrete Problems' distributional-shift section keeps returning to that single hinge.
Counterfactual reasoning and a contracts framing open as descriptive and prescriptive counterparts, before reachability analysis, acting during the search itself, is set against model repair, which waits until training is already finished to intervene.
A cluster of edges in Concrete Problems' distributional-shift section is built from paired framings that all hinge on timing. Counterfactual reasoning is the descriptive counterpart to machine learning with contracts, one diagnosing what shift is, the other prescribing what a system should be built to guarantee; that contracts framing then reframes robustness to distributional shift itself as a specification failure rather than a statistical one. Reachability analysis supplies the safe set that bounds robust policy improvement, acting ex ante, during search; machine learning with contracts is restored post hoc by model repair, and model repair is explicitly the ex-post counterpart to reachability analysis, altering an already-trained model rather than constraining the search that produced it. Read together, the edges show the paper's contracts vocabulary is not decoration: it is what lets "when in the pipeline a fix happens" become a real design choice rather than an afterthought.