This paper proposes a reference architecture for PRETIUM, the internal economics component of the SkipJack Cognitive Exchange suite. The central proposal is deliberately minimal: a non-monetary, non-custodial internal accounting architecture may make heterogeneous AI-capacity economics bounded, ordinally comparable, and auditable by separating six responsibilities—a versioned offer and rate registry; uncertainty-bearing normalization; protected budgets and reserves under closed integer arithmetic; a non-authoritative commitment and proposal lifecycle; evidence-gated clearing with durable unresolved states; and a hard settlement-instruction boundary—without authorizing work, executing capacity, holding value, or declaring any evidence true. The proposal is stated so that it can lose. Three counter-theses receive fair standing. First, normalization futility: uncertainty across unlike meters may be so large that normalized comparison never improves on simple per-provider budgets or an honest not comparable result. Second, boundary instability: any clearing or instruction layer useful enough to deploy may inevitably acquire authorization, custody, or contractual functions, making the proposed hard boundary a paper fiction. Third, no new category: existing FinOps, metering, billing, reservation, chargeback, workflow, and audit capabilities may already satisfy the complete requirement set through composition. If one of those counter-theses survives the proposed evidence program, the paper narrows to specific meter families, a single-channel profile, or a requirement-derived comparison paper, and that narrowing is recorded as a loss rather than concealed. The architecture consists of C1, a registry for versioned offer, rate, meter, freshness, and provenance assertions; C2, a normalization engine that must carry uncertainty and may return not comparable or not distinguishable; C3, a closed internal budget and protected-reserve ledger; C4, a quote, cart, commitment, and non-authoritative proposal lifecycle; C5, usage ingestion and evidence-gated clearing; and C6, a settlement-instruction emitter that accepts only a corrected commitment-level cleared result. Integrity, evidence, and continuity/audit are cross-cutting planes, not sequential components. The edge after C6 is hard: value movement, custody, payment, rail selection, legal obligation, and regulated settlement remain outside. The paper develops this proposal through a hostile-reading vocabulary firewall, bounded dependency assumptions, closed stock and reservation arithmetic, two-layer terminality, a bounded replay and idempotent-effect claim, a nine-family disruptor matrix, capability-class alternatives, a fictional Dynamix conformance case with eleven losable negative controls, and a Pilot-to-Scale evidence ladder. No test has run. No implementation, deployment, market, customer, price, savings, security, compliance, performance, or production claim is made.