Weakening statistical assumptions to survive distribution shift — supplies a route into → Reframing distributional shift as broken contracts and causal structure

The two reframings are not fully independent. The paper's own transition between them names partially specified models, a statistical-relaxations technique, as the first of several routes to the weaker, checkable contracts that machine learning with contracts calls for: partially specified models "offer one approach to this," while "reachability analysis and model repair provide other avenues for obtaining better contracts" (concrete-problems, §"A unifying view: counterfactual reasoning and machine learning with contracts.", p. 19). Covariate shift, the generalized method of moments, and unsupervised risk estimation stay inside classical statistics, treating shift as a matter of which assumption to relax; partially specified models already produces the kind of partial, explicit specification a contract needs. What the contracts framing supplies is not a new technique but an obligation: state that specification openly and check it, rather than silently assume it holds.