Tag
This paper introduces Regime-Conditional Verification (RCV), a lightweight wrapper that adapts off-the-shelf safety classifiers for large language models by estimating prediction correctness and detecting distribution shift without retraining.