<![CDATA[The rapid deployment of reinforcement learning (RL) agents in critical national infrastructure has outpaced the governance instruments designed to hold them accountable, creating a widening gap between algorithmic autonomy and sovereign oversight. This article proposes a Three-Layer Cyber-AI Governance Framework that integrates the technical, organizational, and regulatory dimensions of accountability for RL systems operating within national data sovereignty regimes. Using a systematic literature review guided by PRISMA 2020 procedures, twenty-five peer-reviewed and preprint sources published between 2021 and 2026 were analyzed through thematic synthesis to identify recurring governance constructs across cybersecurity, AI ethics, and data-sovereignty scholarship. The synthesis reveals three interdependent layers: an Algorithmic Layer governing reward design, explainability, and adversarial robustness; an Organizational Layer governing human oversight, audit trails, and incident reporting; and a Sovereign-Regulatory Layer governing data localization, cross-border data flow, and international cooperation. The proposed framework departs from existing layered models by embedding a continuous feedback loop that links real-time algorithmic telemetry to national regulatory review, closing the accountability gap that single-layer or purely technical frameworks leave open. The article concludes that accountable reinforcement learning under conditions of national data sovereignty requires coordinated, multi-layer instruments rather than isolated technical fixes, and it outlines an agenda for empirical validation of the framework across diverse regulatory contexts. ]]>
