Risk-Adaptive Behavioral Attenuation via Episode-Latched Policy Gating and Bounded Relearning for Deployed Reinforcement Learning Agents

Abstract

Reinforcement learning behaviors can become harmful after deployment even when the policy is unchanged, a risk existing methods address mainly before or during training rather than through reversible correction. This paper proposes a risk-adaptive attenuation framework for a frozen context-conditioned policy encoding validated nominal and safe endpoint behaviors. Recurring harm is identified through evidence accumulation, and a severity-aware selector applies gradual, accelerated, or hard-switch attenuation latched for the full episode. Verified probes support bounded relearning, and recurrent harm reactivates the intervention. The selector is a fixed rule, and only the base policy is trained with reinforcement learning. Evaluation uses Risk Corridor, a stochastic-cost gridworld, with ten policies tested across fixed, mixed, and noisy-observation conditions against shield and Recovery RL baselines. Under mixed-severity deployment the selector reached a final hazard rate of 0.00 and task completion of 1.00 at a safety cost of 3.98, above hard switching (2.83) and Recovery RL (0.00). Deviation cost was 44.36, versus 61.20 for Recovery RL (27.5% lower) and 27.81 for the shield, which completed only one of three phases. The framework is best described as a safety–utility trade-off rather than superior safety. Mode-selection accuracy remained above 0.96 at 0.10 observation noise, safety and reversibility held across all tested conditions, and aired differences were consistent for at least nine of ten policies, reaching the exact test’s minimum attainable significance (p=0.00195) for four of five comparators.