The transition from narrow AI to artificial general intelligence (AGI) introduces systemic failure modes absent in task-specific systems: unconstrained optimization convergence, behavioral drift under autonomy, and the elimination of corrective feedback pathways. This paper proposes corrective architecture as a necessary condition for stable AGI-capable systems. We present the HEIMDALL–METIS framework as an empirical implementation: a multi-agent orchestration system operating under a formally immutable security layer with a fixed decision weight of w_security = 0.95. Drawing on the Emergence World simulation study (Emergence AI, 2026), we demonstrate that isolated optimization without corrective feedback produces systemic collapse across all tested frontier models. We argue that immutable corrective weighting, dual-window behavioral baseline monitoring, and hostile-input assumption protocols constitute a generalizable framework for AGI transition stability.