CORE: Uncertainty-Gated Policy Optimization for Reliable Performance Under Distribution Shift

We present Conservative Online Reliability Enforcement (CORE), an uncertainty-gated policy optimization framework for robust policy updates under distribution shift. The method combines risk-aware trust-region learning with online promote/monitor/rollback control driven by uncertainty-calibrated signals. 

Across matched-seed benchmark and scenario-model evaluations, our approach improves reliability-floor behavior while preserving fixed-budget update control. We provide algorithm design, theory-guided gate rationale, and diagnostics that clarify where gated updates help and where limitations remain. 

The public reproducibility package is available online at https://github.com/Aureuma/CORERepro