The Sovereignty Foundation

How certification works

Self-test openly today; certify formally at ratification — one evidence discipline throughout

SF-AGI-PROCESS-001 · v0.1 Public Review Draft

The program runs in two phases with one constant: every claim rides on a sealed, pre-registered, independently reproducible artifact. Phase 1 — open self-testing — needs no one's permission. The corpus and its methodology are published for review. The cells themselves are not yet offered as a runnable suite: the Standard is a Working Draft, it has not been peer reviewed, and no certification has been ratified. Publishing cells against an unratified standard would imply an authority the Foundation has not earned. When the Standard is reviewed and the programme ratified, the cells and their harness will be published without restriction, and any implementer will be able to run them, seal the results, and disclose them publicly. Phase 2 — formal certification — begins when the Standard ratifies, adds the Foundation's independent verification and the certification grant, and follows the Knox-anchored flow below.

Phase 1 — open self-testing


Self-testing follows the same protocol the Foundation's community benchmarking platform, Proving Grounds, already enforces for performance claims — because a conformance claim is a claim, and the protocol exists to keep claims honest:

Claim
Take up a conformance quest for the cell you intend to run — a public signal of active work.
Pre-register
File the hypothesis — which cells, which substrates, expected outcome — before running anything. The registration is immutable and timestamped.
Seal
Run the cells; seal results with the content-addressed manifest and the full hardware context of the run.
Ratify
Independent contributors reproduce on their own hardware and vote. Two ratifications merge the result into the public corpus; one refutation is fatal.

Publishing self-test results before formal certification is encouraged — it is transparency that costs an implementer nothing but honesty, and the merged corpus becomes the public track record that tier migration draws on.

Platform Roadmap Proving Grounds today models performance benchmarks; hosting conformance runs cleanly needs three schema extensions, disclosed here as roadmap: a pass/fail verdict with cell and corpus version fields alongside the latency metrics; evergreen conformance quests that stay open across many implementers rather than closing on first merge; and re-registration support so 24-month re-certification is possible on the same quest.
Where The Evidence Lives — An Honesty Note Proving Grounds is the disclosure and peer-review surface, not the ledger of record. Its own audit view is a derived query over mutable rows — measured against the Standard it would not satisfy AGI-1, AGI-8, or AGI-10. Sealed conformance artifacts therefore anchor in the Foundation's append-only, hash-chained ledger, and the platform links to them. A certification program whose own evidence store failed its own standard would deserve the criticism; this note exists so no one has to discover the distinction the hard way.

Phase 2 — formal certification


At ratification, the certification grant follows the Standard's Knox-anchored flow — six steps, every artifact independently verifiable:

Request
The implementer requests certification at the desired level — Compliant, Certified, or Certified+.
Corpus access
The Foundation issues a capability token authorizing access to the current corpus and harness version.
Sealed run
The implementer runs the corpus and submits results sealed under the implementation's binary hash — the run is bound to the exact code it certifies.
Independent verification
The Foundation verifies the sealed results against the published expected outputs. It does not take the implementer's word; it re-derives.
Grant
A conformance certification capability token issues, and the grant is recorded in the program's public append-only ledger.
Mark use
The implementation may use the certification mark per Foundation licensing terms, at the granted level, until re-certification falls due.

Re-certification runs the same six steps against the then-current corpus version every 24 months. Certification survives PATCH and MINOR standard revisions; a MAJOR revision opens a 24-month re-certification window. Revocations — failed re-certification, the 30-day critical-vulnerability rule, misuse of the mark — are recorded in the same ledger with the same permanence as grants.

Running the scaffolded cells today


Two cells are scaffolded now — tc-agi-1-audit-chain and tc-agi-2-determinism — each a pre-registered cell directory whose harness/RUNNER.md gives the exact test invocations against the reference implementation. "Scaffolded" is used precisely: the scenario is pre-registered and the runner is executable, but inputs are not yet sealed and no manifest.sha256 exists until they are. The remaining eight cells are specified on the corpus page; the five Defense-in-Depth cells are a documented authoring gap.

The Foundation does not take the implementer's word.
It re-derives.