Governing Agentic Work Across Distributed Data and Compute

Agentic work that spans sites, regulatory regimes and days must be governed at every boundary it crosses. This paper follows one synthetic engineering comparison from intent to human acceptance. The comparison joins revisions C and D of a restricted design with restricted test campaigns held at a second site. The paper traces it across six outcome paths: normal, denied, stale, revoked, unavailable and ambiguous. Three author-proposed working definitions explain why the problem is newly hard: Emergent Compute, Agentic Ops and Tokenomics. Their convergence makes each choice at once an authority event, a data-movement event and a cost event. From the drivers and the scenario the paper derives nine problem categories. Delegated identity (5a) and physically segregated administration (5b) are held distinct. The categories are read as the problem of operationalizing Zero Trust for agentic work while preserving human authority, context integrity, data rights and verifiable execution. Two obligations are kept apart: an architectural requirement, and applicable federal mandates, none of which is shown to apply to the synthetic enterprise. The paper then applies the author's HGC³AE² doctrine and introduces aZTF, the author's proposed Automated Zero Trust Framework, with three dimensions: governed lifecycle, automated enforcement and verifiable assurance. Within the proposed Sovereign Data Fabric, the author proposes Legion, a physically segregated administrative plane, as the mechanism for applying approved decisions that take the form of administrative operations on in-scope infrastructure. Resource owners and workflow products keep their own enforcement. A fair composed-gateway counter-thesis is retained at full strength, and every test and falsifier is stated and not yet run. This is the first paper of a nine-paper series. It presents a proposed architecture, with no implementation, test outcome or customer result.