The Capture Exposure Metric: A Testable Framework for AGI Governance Resilience

The governance of advanced artificial intelligence faces a fundamental threat: capture by transnational elite coalitions that use lobbying, regulatory arbitrage, and narrative control to insulate deployment from democratic accountability. This paper introduces a testable governance architecture comprising an append-only AGI Archive, a Fifth Estate of independent auditors, jurisdictional sanctuary counter-measures, and a Capture Exposure Metric (CES) that quantifies institutional vulnerability. We implemented a reference simulator (50 seeds, 20 quarters, Python open source) and executed the Transparency Flooding attack (A2) in two variants:(V1) no archive quality validation; (V2) AQI with self-reported record usefulness. V1 produced a large but directionally inverted effect—CES rose while GTCS rose more slowly, indicating false-alarm injection rather than the predicted camouflage. V2 rendered the attack invisible (DE unchanged, p=0.43), because self-reported metadata under adversarial control carries no independent signal. Neither variant reproduced the theoretical camouflage regime. These findings empirically validate the design paper’s central prescription: an independent semantic oracle for AQI is not optional—it is the minimal condition for the metric to distinguish genuine disclosures from adversarial noise. Sprint 2 will implement the oracle and test suppression attacks