NIMBUS Connect: Governed Execution Across Heterogeneous AI Capacity

Abstract

Organizations that purchase artificial-intelligence capacity increasingly face a boundary problem: an authority layer may decide what work is eligible and issue a bounded grant, and an economic layer may commit a bounded budget against that grant, but the work must still be executed against providers whose interfaces, meters, quotas, localities, and execution modes are heterogeneous and non-fungible. This paper proposes — and does not claim to have implemented or validated — a minimum execution contract for that boundary. The proposed layer, NIMBUS Connect, would consume a bounded grant and a bounded economic commitment; isolate credentials from adapter code paths; enforce retry and failover that cannot widen either the grant or the commitment; correlate every attempt to exactly one authority-and-commitment chain; and return a closed, honest account of success, partial failure, failure, cancellation, expiry, and unresolved knowledge. The paper states the strongest fair counter-thesis: that a documented composition of existing gateways, brokers, identity systems, schedulers, provider meters, tracing frameworks, and endpoint controls may already provide every decision-relevant property, in which case the proposed layer is integration glue rather than a distinct contract. Five conditions under which the proposal loses or narrows are stated explicitly. All contract fields, state semantics, and boundaries herein are proposals rather than observed results.