Auditing Update Admission for Continual Embodied Agents

This paper proposes new metrics and a paired-binomial construction to validate policy updates in continual learning agents. Using a 32-seed diagnostic, the method admitted 31.6% of updates at 2,000 episodes per stage, significantly outperforming standard range-based confidence gates that rejected all updates.