System governance
The machinery that keeps the rules real.
Approved by the Board for implementation, 20 November 2026. Counsel is confirming the open points, shown in [brackets]. This Constitution works under Delaware law, our Certificate of Incorporation and our Bylaws; where they differ, those prevail.
In one paragraph
Under Part VII, every CRIS system would be registered before use and would move through four status levels only as its evidence allows. It would be protected and red-teamed. The People Stewardship Guarantees would apply to every system's data and contributors. Incidents would be contained first, escalated by severity and reviewed. Five layers of assurance, including independent audit and ISO/IEC 42001 certification that CRIS would seek, would back a yearly public report.
The text
Part VII — System Governance
Part VII covers the machinery that keeps the rules real: a register of every system, a lifecycle each must pass through, security, the rights of the people whose data and work feed the systems, incidents, audit and public reporting.
Article 18 — The System Register and Lifecycle
18.1 The Register. The Chief AI Governance Officer shall keep a System Register listing every CRIS System with: its name, version and lineage; Responsible Steward; encoded purpose; registered domains and baselines; autonomy level; Change Class history; deployment locations; status; and open incidents.
18.2 Unregistered means unusable. No CRIS System may be used outside a research sandbox until it is in the Register.
18.3 Status levels. Each system moves through four status levels, and its public status may claim only the level its evidence supports:
| Status | Evidence required | Who decides |
|---|---|---|
| Research | A published concept and draft specification | Chief Executive Officer |
| Prototype | The succession pipeline running on a test model, with ledger records | Chief Executive Officer, on the Labs and Systems leads' advice |
| Pilot | A partner deployment on testnet and an Audit Council review | Chief Executive Officer |
| Live | Production use and an external audit passed | Chief Executive Officer, with notice to the Board |
18.4 Retirement. A system is retired by its Responsible Steward with the Committee's approval. Its records are kept under Article 11.4, and people who relied on it are told in advance where practicable.
Article 19 — Security and Resilience
19.1 Protecting the systems. Model weights, training data, the evidence vault, the gates and the Knowledge Ledger shall be protected by security controls proportionate to their risk, including access control, encryption, separation of duties and monitoring.
19.2 Protecting against misuse. Each system shall be tested before deployment, and at intervals after, against attempts to misuse it, including prompt injection, jailbreaks, data poisoning and model extraction.
19.3 Red-teaming. Systems at A2 and above, and every Class 3 change, shall be red-teamed by people who did not build them, including against the prohibited actions in Article 12.
19.4 Continuity of service. Systems on which people or partners depend shall have tested plans to continue or fail safely if infrastructure, suppliers or networks fail.
Article 20 — Data, Consent and Contributors
20.1 The People Stewardship Guarantees apply. Every CRIS System shall honour the People Stewardship Guarantees in Bylaws §3A.9: consent, attribution, a fair share of value, a human accountable for automated decisions, access to and correction of data, a voice, non-exclusion, and protection that survives a sale of the business.
20.2 Consent for training. No person's contribution, personal data, likeness or voice shall be used to train or test a CRIS System without informed consent or another lawful basis that the Chief AI Governance Officer has approved and recorded.
20.3 Attribution and value. Contributions used by a CRIS System shall be recorded through CIX and valued by the Participation Value Meter under published rules.
20.4 Minimisation. Systems shall use the least personal data their purpose requires, and keep it no longer than needed.
20.5 Explanation and review. A person affected by a decision a CRIS System made or supported shall be able to get a plain-language explanation and to have a person review it.
Article 21 — Incidents
21.1 Severity. Every incident shall be classified by Schedule C, from Severity 1 (critical) to Severity 4 (low).
21.2 Report at once. Anyone who becomes aware of a possible incident shall report it at once to the system's Responsible Steward or the AI incident channel. The Responsible Steward shall contain it and classify it within the time Schedule C sets.
21.3 Escalation. Severity 1 incidents are escalated at once to the Chief AI Governance Officer, the Chief Executive Officer and the chair of the Committee, and to the Board within [24 hours]. Severity 2 incidents reach the Committee within [72 hours].
21.4 Notify people. People and partners materially affected by an incident shall be told what happened, what it means for them and what CRIS is doing, as soon as practicable and in any case within the time the law requires.
21.5 Learn. Every Severity 1 and 2 incident shall have a written review within [30] days: what happened, why, what was changed, and what the Knowledge Ledger now records. Reviews blame causes, not people.
21.6 Regulators. Incidents shall be reported to regulators wherever the law requires, and the Chief AI Governance Officer keeps a register of those duties by country.
Article 22 — Audit and Assurance
22.1 Five layers. Assurance comes in five layers, from fastest to most independent: automated checks (proofs and drift tests); human review (the C-Gate board, every release); independent audit (the Audit Council, every major version); an external standard (third-party certification); and the public report (every year).
22.2 The Audit Council chooses. The auditor, not the team being audited, chooses what to check: every high-risk item and a random share of the rest, recorded before work starts.
22.3 Re-runs. A share of correspondence proofs shall be re-run independently on separate systems. A mismatch beyond tolerance is a finding, and the promotion it supported is reviewed.
22.4 External certification. CRIS shall seek certification of its AI management system to ISO/IEC 42001 by an accredited body, and keep it current, as Bylaws Schedule 1 requires for the Fidelity component.
Article 23 — Transparency and Reporting
23.1 Public reporting. Each year CRIS shall publish, with its public-benefit statement: the System Register (excluding security-sensitive detail); each system's status and autonomy level; summary gate results; incidents by severity; Audit Council findings; and the Fidelity component of the Continuity Function.
23.2 No unearned claims. CRIS shall not describe a system as "verified", "certified", "safe" or "aligned" without naming who verified it and linking to the evidence.
23.3 Specifications open. The Continuity Invariant, the succession pipeline, the gate methods and the audit method shall be published, while model weights, participant data and security details remain protected.
Quoted exactly from the CRIS AI Constitution (version of 8 November 2026, approved by the Board for implementation on 20 November 2026). Headings, lists and tables are only laid out for reading; the diagrams are drawn on this site.
In pictures
- SpotAnyone who becomes aware of a possible incident (21.2)
- Report at onceTo the Responsible Steward or the AI incident channel (21.2)
- ContainThe Responsible Steward contains it first, then investigates
- ClassifySeverity 1 to 4, by Schedule C, within the time it sets
- NotifyPeople and partners materially affected are told what happened and what CRIS is doing (21.4)
- LearnThe review records what happened, why and what changed, in the Knowledge Ledger; it blames causes, not people (21.5)
- Automated checksProofs and drift tests
- Human reviewC-Gate board, every release
- Independent auditAudit Council, major versions
- External standardThird-party certification
- Public reportBenefit report, every year
The same figure appears on the CRL Engine page; here it illustrates Article 22.1 of the proposed AI Constitution.
In practice
For builders. A system outside a research sandbox would first need an entry in the System Register. The entry would record its version and lineage, its Responsible Steward, its purpose, domains and autonomy level. Systems at A2 and above, and every Class 3 change, would be red-teamed by people who did not build them.
For Responsible Stewards. When something goes wrong, the first job is to contain it, then classify it by Schedule C. Severity 1 means stopping the system at once. Serious incidents would get a written review that looks for causes, not someone to blame.
For partners and contributors. Personal data, likeness and voice would need informed consent, or another approved and recorded lawful basis, before training or testing.
Contributions would be recorded through CIX and valued under published rules. A person affected by a system's decision could ask for a plain explanation and a human review. The participant charter sets out what this means for CIX members.
For the public. Each system page has a status block, using the four levels in Article 18.3: Research, Prototype, Pilot and Live. Under the proposed rule, a public status could claim only the level its evidence supports.
Each status is set by a person and links to its evidence. Until one is set, the page says “Status to be confirmed”. See, for example, the C-Gate status.
CRIS would seek ISO/IEC 42001 certification of its AI management system from an accredited body. No such certificate is held today. Under Article 23.2, no system would be called verified, certified, safe or aligned without naming who checked it and linking to the evidence.
The five layers of assurance are the same five shown on the CRL Engine page. They run from automated checks, through human review and independent audit, to an external standard and a yearly public report. Under Article 23.1 that report would be published with the public-benefit statement, as the AI systems report on the benefit report page; the Audit Council's twice-yearly summary under Article 17.5 would be listed there too.
For contributors, the rules on data, consent and contributors in Article 20 are set out at length in the proposed Contributor Rights Charter.
Where it is carried out
- Contributor Rights Charter
- Systems
- Continuity Audit Suite
- Knowledge Ledger
- C-Gate
- Continuity Intelligence Exchange (CIX)
- The People Stewardship Guarantees
- Article IX — AI Governance
