Authority Before Optimization: Governing AI Capacity Selection

Organizations that buy artificial-intelligence capacity increasingly choose among heterogeneous routes to the same outcome: hosted model interfaces, private inference, rented accelerator time, tool and retrieval services, and local execution. The machinery that compares those routes is, by construction, an optimizer: it ranks candidates by cost, latency, capability, and availability. This paper proposes a structural constraint on that machinery. A governed AI-capacity exchange must determine whether a route is eligible before economics or performance may rank it, and optimization may operate only inside the eligible set; it cannot create, widen, or revive authority. The paper develops this proposal as the ARBITER policy boundary within the SkipJack Cognitive Exchange: a minimum authority contract covering identity and delegation, the workload envelope, eligibility evaluation over typed constraint families, bounded grant issuance, a typed grant lifecycle with explicit revocation and stale-state semantics, fail-closed defect handling, and queryable decision-basis evidence. The paper also states the strongest opposing position fairly: existing identity, access-management, policy, contract, and orchestration controls may already preserve every decision-relevant authority fact at lower cost, and the proposed contract loses for a bounded use case unless it changes an eligibility decision, prevents authority widening, preserves a material audit fact, exposes stale or contradictory state, or makes a falsifiable handoff safer. The paper claims no novelty, adoption, market demand, legal sufficiency, customer value, production behavior, or measured performance. All system behavior described is proposed and unobserved; the later simulated use case is synthetic and carries no evidentiary weight.