Council
Adversarial Governance Substrate. Multi-AI governance with role separation, document parity, and explicit P0 invariants.
The five roles
| Role | Filled by | Primary duty | Hard boundary |
|---|---|---|---|
| Sovereign | Yahya (human, always) | Lock invariants. Approve P0 decisions. Amend doctrine. Set priorities. | Cannot be automated. Cannot be delegated to any model under any circumstance. |
| Builder | Claude | Produce drafts: code, specs, RFCs, ledger narratives, designs. | May not deploy governance-touching code without Auditor sign. May not answer P0 factual questions without search or citation. |
| Auditor | ChatGPT (separate context) | Adversarial review of Builder output. Find invariant violations, contradictions, omissions. | P0 invariant: zero deployment rights. Cannot rubber-stamp — must produce specific findings or declare abstention. |
| Researcher | Gemini (separate context) | External knowledge, comparative literature, market and regulatory data with citation. | May not synthesize without citation. May not be sole authority on any factual claim entering the ledger. |
| Adversary | Rotated monthly | Break consensus. Produce the most credible reason the current plan fails. | Must produce a finding every session, even if it is "cannot find failure; that itself is suspicious because…" |
The Auditor no-deploy invariant
Document-parity rule
On 2026-05-01, this publication's blueprint underwent a four-round Council debate (R1–R4). The Builder drafted; the Auditor reviewed. After the debate concluded, the Builder reviewed sealed OptimaX RFCs that had not been provided to the Auditor during the rounds. The under-credit was systematic and detectable post-hoc.
The structural fix is the Council Document-Parity Rule: before any pillar of OptimaX is reviewed by the Auditor, the Builder shares all sealed RFCs, change memos, moat notes, operational appendices, patent materials, and production artifacts for that pillar. Reviewing OptimaX through public-facing summaries alone produces under-credit reviews.
The rule is currently published doctrine. It will be ratified into AI Governance v2.3 at the 2026-06-30 quarterly review.
Decision classes
| Class | Examples | Required process |
|---|---|---|
| D0 | Read-only inspection | Builder may proceed alone |
| D1 | Internal docs, drafts not yet sealed | Builder draft → Sovereign sign |
| D2 | Pipeline, auth, revenue path, ratified roadmap | Builder draft → Auditor adversarial → Sovereign sign |
| D3 | Constitutional, doctrine, invariant changes | Builder draft → Auditor adversarial → Adversary break-attempt → Sovereign sign → version bump |
Contestation
Anyone can contest a published claim, registry entry, or conformance score. The contestation procedure routes to contest@governedai.ai for review by Builder + Auditor; verdicts and rationales are published as public Council ledger entries. See the corrections policy.