Structural Capability Containment in Advanced AI Systems A Foundational Framework for Multi-Jurisdiction Safety Architectures

Abstract

<p><em><span>Advanced AI systems risk accumulating unified decisional authority as capability scales. This paper proposes a structural safety paradigm: instead of ensuring models choose good actions,<b> we design  systems where no component possesses unrestricted authority</b>. </span></em></p> <p><em><span>We formalize jurisdictional architecture, where actions require cross-domain authorization from independent subsystems (policy, impact, task, stability). Under uncertainty, legitimacy thresholds prevent default to “least bad” optimization, <b>introducing structural inertia as a safety mechanism</b>. </span></em></p> <p><em><span>We demonstrate that <b>catastrophic risk reduces multiplicatively with jurisdictional separation</b>, and that increasing model capability does not automatically increase allowed impact. This foundational framework <b>decouples intelligence from authority</b>, enabling safe scaling of advanced AI through bounded institutional architecture.</span></em><span></span></p>