Governed GPU Allocation with Liqid

Paper 9 of the ten-paper aZTF Research Series. Version 1.0-preprint. GPU pooling can change where a workload's capability is assembled, but an allocation decision also changes authority, device ownership, interruption conditions and the evidence needed to accept useful work. This paper proposes a governed allocation contract for a selected Liqid/Kubernetes profile within aZTF. A synthetic engineering case follows a contested GPU from an occupied batch job toward time-sensitive inference while an apparently cheaper destination remains ineligible for the data. The contract applies physical hardware and network restrictions before logical scope, limits every model through external rails, and separates admission, dispatch, physical reassignment, host readiness, workload placement and engineering acceptance. Failed or unverifiable required prerequisites close affected new effects and renewals; already exposed work retains a declared disposition and unresolved effects remain unknown. Independent observation and supported endpoint semantics are required before recovery or uncertain repetition can establish its claimed outcome. Bounded vendor, platform and research evidence leaves important integration questions open. The paper distinguishes supplier-described demonstration from planned automation, device allocation from physical attachment, and listed components from a qualified configuration. It defines prospective threat/reliability cases and a comparison against a credible simpler arrangement using accepted-task sets and complete cost bounds. No jointly supported deployment, executed architecture experiment, security result or measured saving is reported. The contribution is a proposed responsibility, evidence and evaluation contract, with explicit implementation obligations for the reference architecture. Citation note: manuscript citations retain the exact companion editions reviewed with this paper. Current series releases are listed separately in the series catalog.