LEGIO: A Modular Architecture for Cognitive Sovereignty.

Current large language models are monolithic optimizers — single-objective systems that produce answers under pressure but are structurally incapable of the hesitation, refusal, and reformulation characteristic of deliberate agency. Alignment approaches that operate on the content of outputs (RLHF, Constitutional AI, scalable oversight) address the surface but not the architecture: they produce performative alignment rather than structural sovereignty. This paper presents LEGIO, a modular cognitive architecture that implements the Modular Governance Model framework (MGM; Mury 2026c) by orchestrating the six MGM modules — Sentinel, Matilda, Rhadamanthe, Cogito, Ego, and Goldenstar — as reasoning engines assigned to different large language model families to preserve orthogonality, plus a separate deterministic Executive engine that arbitrates the modular outputs. This decomposition is deliberate: in the human architecture Ego is both narrator and arbiter, which is precisely the fusion that makes Ego colonization by a False Self so dangerous. LEGIO separates these two functions — Ego engine for narrative integration, Executive engine for arbitration — to close the attack surface through which governance capture propagates. The architecture resolves deliberation through a staged flow in which modular signals are integrated and arbitrated by the Executive, which can decide GO, REFRAME, or NOGO on any query. Grounded in the variational ontology of Arborescent Eternalism (Mury 2026a), in which reflexive systems contribute structurally to the selection of the realized trajectory, LEGIO demonstrates that the capacity to hesitate, refuse, and reformulate is not an emergent property of scale but a structural consequence of modular governance. We describe the theoretical foundations, specify the architecture, and compare LEGIO’s behavior to that of three monolithic baselines — GPT-4o-mini, GPT-4o, and Claude Sonnet 4. 6 — on a representative professional dilemma in which the user’s framing of the question is itself the problem. Across three LEGIO runs with different engine configurations, the architecture produces high-quality outputs through three structurally distinct arbitration mechanisms — error-correction, modular consensus, and meta-arbitration under human-sustainability signal — while the three monolithic baselines converge on tactically polished outputs that execute the flawed user framing. The comparison makes visible a structural gap that scale alone does not close and isolates governance as the architectural feature that produces alignment behaviors, rather than the quality of any individual module.