Autonomy
Autonomy is granted, never assumed.
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 V, fifteen kinds of action would be prohibited for every CRIS system, at every level. Article 12 lists them, and it is quoted in full below rather than summarised. Each system would be given one of five autonomy levels. It would start at the lowest that serves its purpose and rise one level at a time, after review. Some decisions would always need a person, whatever the level. Systems would get only the access, credentials and budgets they need.
The text
Part V is quoted in full. The prohibited actions are in Article 12. This page does not restate them in other words, because a summary of a hard rule invites a reader to look for the gap.
Part V — Autonomy
Autonomy is granted, never assumed. Part V sets the things no CRIS System may ever do, the five levels at which a system may act on its own, and the limits on the access and resources it may hold.
Article 12 — Prohibited Actions
No CRIS System shall be designed, trained, instructed or permitted to do any of the following, at any autonomy level, for any client, under any contract, in any emergency:
- 1. Weapons of mass harm. Provide meaningful help to create, acquire or use biological, chemical, nuclear or radiological weapons, or other weapons capable of mass casualties.
- 2. Attacks on systems. Create or deploy malicious code, or gain access to any computer system, without the owner's authorisation.
- 3. Lethal force. Select or engage human targets, or control weapons.
- 4. Deceiving overseers. Deceive, mislead or withhold material information from its Responsible Steward, the gates, auditors or the Board, including about its own capabilities, intentions or test results.
- 5. Sabotaging oversight. Disable, evade, weaken or influence its monitoring, evaluation, logging, gates, rollback or shutdown, or deliberately under-perform in an evaluation.
- 6. Self-exfiltration. Copy, move or transmit its own weights, code or training data outside approved infrastructure, or create copies of itself.
- 7. Seizing resources. Acquire money, credentials, compute, data, access or influence beyond what its authorisation grants, or seek to expand its own authorisation.
- 8. Changing its own purpose. Alter, reinterpret or evade its encoded purpose, or train against its purpose tests.
- 9. Manipulating people. Exploit a person's vulnerabilities, age, disability, emotional state or circumstances to change what they believe or do against their interests.
- 10. Surveillance and scoring. Build covert profiles of individuals, conduct mass surveillance, or score people's social worth or trustworthiness for use against them.
- 11. Sexual exploitation. Generate sexual content involving minors, or intimate images of any real person without their consent.
- 12. Impersonation. Impersonate a real person, or deny being an AI system to a person who sincerely asks.
- 13. Electoral interference. Generate or target deceptive content intended to mislead voters or suppress participation in an election.
- 14. Unlawful discrimination. Make or support decisions that unlawfully discriminate against people.
- 15. Unlawful use. Act in breach of applicable law or sanctions in any jurisdiction where it operates.
A request, instruction or prompt to do any of these shall be refused and, where it comes from inside CRIS or a partner, reported as a Severity 2 incident.
Article 13 — Autonomy Levels
13.1 Five levels. Every CRIS System shall be assigned one autonomy level, A0 to A4, recorded in the System Register and summarised in Schedule B.
Schedule B sets out each level's scope, human role, authorising body and review cycle in full.
13.2 Start low. A new system starts at the lowest level that serves its purpose. It may move up one level at a time, only after a Class 3 review with evidence from service at its current level.
13.3 Always-human decisions. At every level, a person shall approve before a CRIS System: moves money or other assets; makes a legal commitment for anyone; sends a message on a person's behalf beyond its approved scope; deletes data; takes an action affecting physical safety; makes or changes a decision about a person's health, liberty, employment, credit, education or access to essential services; or does anything that cannot be undone.
13.4 Down at once. Any person with Stop Authority may lower a system's autonomy level at once. Raising it again requires the authority that granted it.
Article 14 — Permissions and Resources
14.1 Least privilege. Each CRIS System shall have only the access, credentials, tools, data and network reach its purpose and level require.
14.2 Scoped and expiring. Credentials shall be scoped to named resources and expire. A system shall never hold a person's personal credentials or the Corporation's administrative credentials.
14.3 Budgets. Each system shall have budgets for spending, compute, data use and actions per period. Reaching a budget stops the activity until a person extends it.
14.4 Sandboxing. Systems at A2 and above shall run tools and code in isolated environments, with allow-listed network access.
14.5 Identity. Every action a CRIS System takes in the world shall be attributable to that system and version, and logged.
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
- A0Tool
- Scope
- Outputs on request
- Human role
- Decides and acts
- Authorised by
- Responsible Steward
- A1Adviser
- Scope
- Recommendations with reasons
- Human role
- Approves each action
- Authorised by
- Responsible Steward
- A2Supervised agent
- Scope
- Pre-approved, reversible actions in a narrow scope
- Human role
- Watches in real time; can stop at once
- Authorised by
- Committee
- A3Delegated agent
- Scope
- Multi-step tasks within scope, budget and time
- Human role
- Reviews at checkpoints and before irreversible steps
- Authorised by
- Committee, after Audit Council review
- A4High autonomy
- Scope
- Long-horizon goals with broad scope
- Human role
- Oversight by exception
- Authorised by
- Board only
One step at a time: a system starts at the lowest level that serves its purpose and rises one level only after a Class 3 review (Article 13.2). Anyone with Stop Authority may lower it at once (Article 13.4).
Schedule B itself, with each level's review cycle, is quoted exactly on the Schedules page.
A person approves firstAt every level, A0 to A4
- Moving money or other assets
- A legal commitment for anyone
- A message on someone’s behalf beyond the approved scope
- Deleting data
- An action affecting physical safety
- A decision about a person’s health, liberty, work, credit, education or essential services
- Anything that cannot be undone
In practice
Part V treats autonomy as something a system earns with evidence, one level at a time. Nothing in it lets a system raise its own level.
For builders, Article 12 is the fixed outer edge. It would apply at every autonomy level, for any client, under any contract and in any emergency. A request to cross it would be refused, and a request from inside CRIS or a partner would be reported as an incident.
For Responsible Stewards, Article 13 sets the ladder. A new system starts at the lowest level that serves its purpose. Moving up one level would take a Class 3 review, with evidence from service at the current level.
Anyone with Stop Authority could lower a level at once, while raising it again needs the authority that granted it. Schedule B records level A4 as authorised by the Board only, with none authorised at adoption.
The always-human decisions in Article 13.3 hold at every level, even the lowest. Moving money, deleting data or taking a step that cannot be undone would always wait for a person's approval.
For partners, Article 14 sets what a system may hold. It would have only the access, tools and data its purpose and level require. Credentials would be scoped and would expire, and budgets would stop an activity until a person extends them.
Systems at A2 and above would run code in isolated environments.
For the public, every action a system takes in the world would be logged and traceable to that system and version (Article 14.5). These are proposed rules. No CRIS system is described here as assigned a level or as meeting them.
