The problem as the field states it
Nadella's formulation is the compact one: there is no global alignment architecture for AI, and one is needed. The observation generalizes beyond alignment. Every regime governing what may be computed on what data — privacy law, sectoral regulation, contractual restriction, consent, professional obligation — is enforced administratively, after the fact, within a jurisdiction, by inspection. The result is that the most valuable data is the data a model may not touch, and that AI operates lawfully only in the shallow end of the economy.
Why the substrate made it unsolvable
Compliance became a cost center because it was bolted on. A model does not know what it may do; a compliance function determines that separately, per jurisdiction and per counterparty, and enforcement depends on audit rather than on architecture. Building it in required governance to be a property that travels with the resource through unlimited recombination, which was the thing that could not be done.
What the paradigm supplies
Universal Compliance makes lawfulness a property of the computation rather than a review of it. Every computation is bound to verifiable rights, every agent constrained by lawful purpose, every action auditable and reproducible, every resource gated by trust rather than by capability, and every autonomous system dependent on human sponsorship. Verification is continuous and cryptographic, operating without centralized inspection, proprietary disclosure or coordinated regulatory adjudication across jurisdictions.
The consequence for AI is that the constraint inverts. Regulated, proprietary and personal data becomes usable for computation, orchestration, personalization and inference at global scale, with every participant's rights continuously enforced. The binding constraint on AI was never capability; it was permission. This is the mechanism that supplies permission, and it is what makes the first candidate in this part deployable rather than theoretical — containment is only meaningful in a system where the agent's resource access is governed in the first place.
A demonstration that a contained agent can formulate a prohibited action, that revocation paths are not independent, or that verification requires disclosure after all.