The Sovereignty Foundation
The AGI Certification Program
A certification program that is itself certified
The certification pipeline: claim → pre-register → seal → ratify → countersign. The decision is produced by the same sealed, peer-ratified evidence discipline it certifies.
Artificial Governed Intelligence redefines the category: an AI system whose governance, audit, and authorization properties are constitutive of its capability rather than bolted on. The Standard specifies that claim as eleven testable requirements. The certification program makes the claim checkable by anyone — an implementation passes the published Conformance Test Corpus or it does not, and there is no policy negotiation around the bar.
What makes this program unusual is that it runs on the discipline it certifies. The standard's requirements are named test cells. Conformance evidence is sealed, content-addressed, and reproducible — though no cell has yet been independently reproduced. The certification decision itself is a ratified, ledger-recorded event. One evidentiary primitive, applied at every altitude — including to the Foundation's own reference implementation, whose shortfalls are published in the Known Deltas register before certification of anyone is offered.
Why this exists
In May 2025 the Foundation's founder published the argument this program rests on — “Toward Artificial Governed Intelligence: A Moral Imperative in the Age of AGI”: that the defining question of advanced AI is not generality but governance — that a system whose governance can be stripped away was never governed at all, and that AGI should therefore mean Artificial Governed Intelligence. It was the first thesis of a ten-essay corpus published that month — including the ontological restatement “AGI = Governed Intelligence” and a human-in-the-loop white paper that already named RITA, SAGE, and Mirror. At the time it was argument. In the fifteen months since, it became a working system: an architecture built in the discipline of top-tier high-assurance design — separated trust domains, cryptographically gated execution, append-only evidence — reduced to a filed patent estate and a running reference implementation whose test counts appear in the footers of the Standard itself — the full dated chain is laid out in the evidence record. What distinguishes it from the architectures it mirrors is not a claim of superiority; it is that the evidence is published, reproducible, and open to refutation.
The institutional form is not incidental — it is the qualification. A certification authority's judgments cannot be downstream of quarterly performance, which is why the Foundation is a nonprofit standards body holding its intellectual property in custodial stewardship, with operating companies structurally beneath it and governance authority flowing only downward. That neutrality is what a certification mark requires to mean anything, and it is what lets the Foundation engage governments, institutions, and private companies as a standards body rather than as a vendor.
The Standard is now being piloted in production — in the reference implementation and in the infrastructure of the Foundation's first pilot cohort: DexaFit, Cracker Digital Marketing, and Pulse Mesh, operating companies with which the Foundation shares ownership or control, with further pilots to follow. The pilots run the requirements under real load and publish self-test evidence like any Early Adopter. They hold no marks and never will: the certifier does not certify itself, and it does not certify its affiliates — the cohort's role is to be the evidence, not the endorsement.
The architecture it certifies
The four trust boundaries of AGI-9 — Evaluator, Minter, Enforcement Gate, Audit Ledger — each with distinct key material, communicating only through independently verifiable artifacts. A newly composed figure, not a reproduction of any filed drawing.
Compromise of any single boundary is structurally insufficient to forge the others' artifacts. The requirements run from the audit chain every governance decision lands in (AGI-1) to the thermodynamic cost that makes mass authorization forgery economically infeasible (AGI-11) — each anchored, in the Standard's own text, to the filed patent estate and to the reference implementation's code and test counts.
Versioned honestly
Four requirements the Foundation can defend with working code today: the audit chain, fail-closed enforcement, the append-only ledger, the multi-trust-boundary architecture.
Five requirements with working evidence whose formulations invite community refinement before hardening to MUST in v2.0.
Two research-active requirements included to flag the territory and invite contribution — privacy-preserving attestation and thermodynamic authorization cost.
Requirements migrate between tiers on implementation evidence, not preference — and the ratchet clause means the bar only moves up. The corpus applies the same honesty to itself: fourteen at reference pass, one partial, two scaffolded, eight specified, five deferred — none independently reproduced, stated plainly on every page that mentions them.
The program documents
Governance is constitutive of capability.
Certification makes the claim checkable.