Modular Governance for Hybrid Human-AI Agency: A Cognitive Architecture for Shared Arbitration Over Time — System Demonstration

Current AI systems optimize user queries even when structurally incoherent, producing four compounding failures: sycophantic drift, scalar collapse of plural normative constraints, narrative drift, and accountability gaps. We present LEGIO — a computational cognitive architecture that orchestrates several functionally specialized reasoning engines, each implemented by a different LLM family to preserve orthogonality, with per-engine deterministic constraints. A hybrid Executive Engine produces GO, REFRAME, or NO-GO verdicts. The demonstration walks through five independent cases and compares the outcome of LEGIO versus a fixed monolithic baseline (GPT-4o). In each case, LEGIO illustrates that the central challenge of hybrid human-AI agency is not output quality but decision governance.