Model repair — is the ex post counterpart to → Reachability analysis

The paper lists these together as two ways of obtaining a "better contract" on a system's behavior (concrete-problems, §"A unifying view: counterfactual reasoning and machine learning with contracts.", p. 19), but they intervene at different times. Reachability analysis is ex-ante: it constrains the optimization or policy-search process itself, by requiring that a safe region remain reachable via some known conservative policy throughout. Model repair is ex-post: it takes a model that has already finished training, discovers a desired safety property does not hold, and alters the model directly to force it to hold, without redoing the training that produced the violation in the first place. The choice between them is a choice about when in the pipeline the safety property gets enforced — during search, or after the fact once a problem has already been found.