Selected-Risk Calibration (SRC) for surrogate-feasibility guarantees at returned decisions
Selected-Risk Calibration (SRC) is a framework for calibrating a learned surrogate constraint using only the optimizer’s returned decisions. Instead of guaranteeing coverage over random inputs, SRC targets the probability that a decision produced by the surrogate-based optimizer is actually infeasible under the true constraint, for objectives drawn from a specified distribution.
Discovered 2026-11-10 · Sources checked 2026-11-10
Why it matters
Surrogate constraints can label a decision as feasible even when it violates the true constraint. SRC addresses this mismatch by calibrating to the subset of outcomes induced by the optimizer’s selection mechanism, yielding a controlled “selected risk” for returned decisions.
What changed
SRC redefines calibration from input-coverage notions (e.g., split conformal coverage) to risk assessed on the optimizer’s selected/returned decisions. It also accounts for situations where reweighting methods for feedback covariate shift do not apply because the learned constraint is active at the returned decision.
Where it helps
- learning constraints for optimization under surrogate models
- certifying feasibility of decisions produced by a learned constraint optimizer
- post-training calibration of constraint models using returned-decision feedback
- settings where true constraint violations must be bounded for selected outcomes
Evidence & reproduction
Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.
Limitations & open questions
The guarantees and applicability depend on having feasibility verifications for the optimizer’s returned decisions on sampled instances/objectives drawn from the given distribution; SRC does not provide coverage-style guarantees for random inputs in place of selected-risk guarantees.
Related World Models papers
Sources
- Selected-Risk Calibration for Feasibility at Returned Decisions ↗
arXiv · primary · Checked 2026-11-10
