Can uncertainty estimates stay fair when data shifts over time?

Real-world ML systems must provide reliable uncertainty estimates that don't degrade as data distribution shifts—and crucially, those estimates should be fair across different demographic groups. This paper combines two usually-conflicting goals: group-conditional coverage (fairness) and parameter-free adaptation (robustness to unknown shifts). The algorithm achieves tighter prediction intervals than existing tuned methods while guaranteeing coverage for every subgroup, tested on synthetic and real datasets.