Split-Authority AI Coordination

AI-assisted construction often lets one conversational agent interpret a request, choose an architecture, write artifacts, run tests, judge success, and present the result. This concentration is efficient, but it hides where intent became specification, where specification became implementation, who authorized deviations, which evidence supports success, and whether an external effect occurred. A coherent output can therefore validate the assumptions that produced it. The source paper introduces Split-Authority AI Coordination (SAAC), using Governor, Coder, Codex, and Journal roles to separate semantic authority from execution. This revision preserves that insight while expanding the protocol into five authority-bearing functions: mandate, specification stewardship, execution, independent verification, and external effect. The Codex is a versioned contract binding these functions; the Journal becomes a custody-preserving evidence and correction ledger rather than an informal feedback channel. The central correction is that role separation does not itself create legitimate authority or correctness. A Governor receives a bounded mandate; it does not own the user's meaning. A Coder necessarily interprets during lowering and must surface gaps rather than pretend to execute without interpretation. A verifier can establish only the claims supported by its observations and tests. An effect authority can commit only within a named capability and jurisdiction. Logical roles can still collude, share correlated failures, or be instantiated by one model under multiple prompts. Independence, non-collusion, source custody, claim ceilings, and recovery must be tested. SAAC therefore makes a narrower claim than the source. Under an exact Codex, bounded role authorities, independently checkable evidence, declared variability, and explicit effect gates, split authority can make specification drift and self-validation more visible and easier to refuse. It does not guarantee identical outputs, specification completeness, correctness, legitimacy, or cross-model reproducibility. The paper supplies a formal transition protocol, conflict and deadlock rules, cold-boot and provider-substitution obligations, counterevidence, conformance requirements, and placement in the full realization funnel.