Organizations that buy AI capacity now purchase across model APIs, hosted inference, rented and preemptible GPU capacity, tool and retrieval services, and private local execution — capacity classes whose native meters do not share a unit, whose prices do not encode eligibility, and whose failures do not respect budget lines. This paper proposes, and does not claim to have implemented, a distinct governed economic contract for that purchasing problem: PRETIUM, an economic layer that operates strictly inside an externally owned authority boundary. The proposal has four elements: normalization that permits bounded comparison of unlike capacity without asserting fungibility; protected internal reserves for verification, synthesis, recovery, and closeout obligations; commitments bound to an eligible set that economics cannot widen; and evidence-backed clearing over an explicit lifecycle whose arithmetic must balance. The paper states the strongest counter-thesis fairly: that existing budgeting, reservation, billing-export, and dispute mechanisms — FinOps plus escrow — may already preserve every decision-relevant economic fact at lower cost. The proposed contract is offered with precommitted loss conditions: it loses if it cannot change a bounded eligible-set decision, protect a completion obligation, expose a material uncertainty or failure state, preserve a reconciliation fact, or outperform a declared naive cheapest-eligible policy under a test fixed before evaluation. No novelty, adoption, demand, legal status, financial value, production behavior, or measured performance is claimed.