Structural Self-Governance as a Criterion for Artificial General Intelligence

We argue that “AGI” requires more than broad task performance: it requires structural self-governance—the ability of a deployed system to modify its own internal structure while (i) preserving safety-critical competence under adversarial distribution shift, (ii) preserving worst-slice general reasoning performance, and (iii) remaining within bounded resource budgets under non-stationary workloads. We contribute (1) a binding, audit-grade operational standard for autonomous structural change, the Standard for Autonomous Structural Management (SASM Rev. 1.5); (2) a reference architecture, the Recursive Causal Synthesis Agent (RCSA), that implements SASM via deterministic acceptance gates, mandated adversarial replay in shadow evaluation, and deterministic rollback; and (3) diagnostic experiments (including ablations and Monte Carlo designs) demonstrating two fundamental failure modes in self-modifying systems under pressure: self-blinding (trading safety for cheap resource relief) and virtuous rigidity (preserving safety but accumulating runaway debt without synthesis). Finally, we provide a Verification & Validation (V&V) plan, a requirements traceability matrix, and an evidence-bundle schema enabling peer reproduction and third-party audit.