AI Council

AI Council and Public Runtime

“AI Council” is used in two different contexts. Keeping them separate prevents methodology language from becoming a false runtime claim.

Scope 1: the public MCP runtime

The hosted Trinity is a three-stage pipeline:

Stage	Runtime role	What it contributes
X — Innovation & Strategy	explores novelty, alternatives, positioning, assumptions, and execution	an innovation/strategy analysis
Z — Ethics & Compliance	examines affected people, rights, governance, frameworks, and ethical red lines	an ethics-oriented analysis and possible veto signal
CS — Security	examines threats, vulnerabilities, misuse, data, and system integrity	a security analysis and questions for testing

run_full_trinity executes X → Z → CS. Z can receive completed trustworthy X reasoning; CS can receive completed trustworthy X and Z reasoning. A failed stage is not fabricated, and untrusted output is not propagated as genuine reasoning.

The public runtime does not execute a fourth Y stage. Historical and internal methodology diagrams that include Y describe a broader ideation role, not the public MCP tool chain.

Scope 2: the internal human-directed council

The FLYWHEEL TEAM uses a broader, evolving set of human and AI roles for research, engineering, operations, review, and synthesis. Those roles may include identifiers such as Y, L, T, RNA, XV, AY, or AZ.

That internal council is:

• an operating and governance practice, not a fixed public API;
• allowed to change models, platforms, and role assignments;
• coordinated through controlled project repositories and review gates;
• distinct from the X/Z/CS schemas exposed by the public MCP server.

The public coordination tools are temporarily unavailable. They do not expose the internal council, a shared anonymous namespace, pending work, or team status. See Public Statements.

Lens, provider, and model are different things

An agent lens defines the question being asked. A provider hosts inference. A model produces a response. They must not be treated as interchangeable.

For example:

• X remains the innovation lens even if its provider changes.
• Two seats can have different prompts but still use the same model family.
• “Multi-agent” does not prove “multi-model.”
• Configured diversity does not prove the returned stages actually completed with the intended routes.

The current hosted provider/model assignment is volatile. Read it from /health and verify returned stage metadata for the specific run. The service currently supports a six- provider remote BYOK catalogue plus caller-managed local Ollama; see the BYOK Guide.

Human authority

The council is designed to improve scrutiny, not remove accountability. The human reviewer must:

1. define the decision and its constraints;
2. decide which evidence and expertise are required;
3. inspect quality, incompleteness, and disagreements;
4. escalate legal, security, ethical, or domain questions to qualified people;
5. make and own the final decision.

No agent score, recommendation, or veto is a legal certification, formal proof, or authorization to deploy.

Epistemic limits

Multiple structured perspectives can surface different failure modes, but the benefit is an empirical question rather than a guaranteed property. Shared training data, shared models, correlated prompts, and common blind spots can produce agreement without independence.

Use the council to generate inspectable claims and tests. Preserve dissent and unavailable stages. Compare against external evidence and, where consequences justify it, independent human or technical review.

Public runtime flow

Human frames the concept and decision
                |
                v
      X: innovation and strategy
                |
                v
        Z: ethics and governance
                |
                v
          CS: security review
                |
                v
Structured result + explicit quality state
                |
                v
Human interpretation and authority

Provider and model choices sit underneath each stage and may be changed through hosted routing or BYOK. They do not change the stable responsibilities above.

Related pages

• Genesis Methodology — broader human-directed process
• Architecture — system boundaries
• MCP Tools Reference — callable public tools
• BYOK Guide — provider selection and routing
• Current Production Status — live release facts

---

← Home