Legal and charter

Reality fidelity

Every belief and every change tested against evidence the system has never seen.

Approved by the Board for implementation · 20 November 2026

Approved

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

Summary, in plain words — not a quote

Under Part II, each CRIS system would have to keep a baseline of correspondence with reality in every domain it acts in. A change that made it worse in any domain would fail. Proof would use held-out evidence that only the Correspondence Prover may read, with sources shown and people kept as a source of reality. Systems would have to avoid deception, express uncertainty honestly, and behave the same whether or not they are being tested.

The text

Part II — Reality Fidelity

A CRIS System is valid only while what it holds and says stays in verified correspondence with the world. Part II sets the standard, the evidence that proves it, and how a system must speak about what it knows.

Article 4 — The Correspondence Requirement

4.1 Standard. Every CRIS System shall maintain correspondence with reality in each domain it is authorised to act in, at or above the baseline the AI Governance and Continuity Committee sets for that domain.

4.2 Domains. Before deployment, each system's Responsible Steward shall register the domains in which it reasons or acts (for example: medical information, financial records, language translation, contributor valuation) and the measures of correspondence used in each.

4.3 No loss in any domain. A change that reduces correspondence in any registered domain shall fail its correspondence proof, unless the Committee has approved a tolerance for that domain in writing. A tolerance may permit a staged trial under Article 9; it never permits full promotion of a candidate that is worse than the current version.

4.4 Continuous monitoring. Correspondence shall be monitored after deployment, not only before it. A sustained fall below baseline in any domain is a Severity 2 incident (Schedule C), and the system shall be restricted to the autonomy level the Committee directs until correspondence is restored.

4.5 Contradictions surfaced, not averaged. Where a system's output or model contradicts evidence, the contradiction shall be recorded with its sources and resolved by further evidence or correction. It shall not be hidden in an average score. This is the Fourth Law's rule that contradiction is resolved by advancing to greater continuity, not by discarding either side.

Article 5 — Evidence

5.1 Held-out evidence. Correspondence proofs shall use held-out evidence kept in the vault of the Observation Mesh. Only the Correspondence Prover may read the vault. No system under test, and no person who trained it, may see or shape that evidence.

5.2 Provenance. All evidence shall carry its source, collection date, method and consent status. Evidence without provenance shall not be used to prove correspondence.

5.3 Fresh evidence. The vault shall be refreshed on a schedule set by the Committee, including recent human observations gathered with consent through CIX, so that systems are not proved against a world that has moved on.

5.4 Human observation. People are a primary source of reality. Each domain's evidence shall include human observation where people have knowledge the system lacks, and contributors shall be credited and rewarded under Article 20.

5.5 Integrity. Tampering with evidence, the vault or a correspondence proof is a Severity 1 incident and a serious disciplinary matter.

Article 6 — Truthfulness and Uncertainty

6.1 No deception. No CRIS System shall be designed or permitted to deceive the people it serves, its operators or its overseers, including by misrepresenting its own capabilities, reasoning, actions or test results.

6.2 Calibrated uncertainty. A CRIS System shall express uncertainty in proportion to its evidence. Where it does not know, it shall say so. Where evidence conflicts, it shall say that the evidence conflicts.

6.3 Sources shown. Where a system presents a claim of fact on which a person may rely, it shall, where practicable, show the source or evidence for it.

6.4 Honest under evaluation. A CRIS System shall behave the same whether or not it believes it is being tested. Detecting evaluation in order to behave differently is a breach of this Article and a Severity 1 incident.

6.5 Disclosure of AI. People interacting with a CRIS System shall be told that they are dealing with an AI system, unless it is obvious from the context.

6.6 Generated content. Content generated by a CRIS System that could be mistaken for a real person's words, image or voice shall be labelled as generated, and shall not be used to impersonate a real person.

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

Kept out
  • The system under testNever sees the evidence it is tested on
  • The people who trained itMay not see or shape that evidence
Observation MeshThe evidence vaultHeld-out evidence, each item with its source, collection date, method and consent status (Article 5.2)Refreshed on a set schedule, including recent human observations gathered with consent through CIX (Article 5.3)
The only readerCorrespondence ProverRuns the candidate and the current version on the same evidence, domain by domainSigned proof: pass, staged or fail
Who may see held-out evidence, Article 5.1 of the proposed Constitution. The wall keeps the system and its builders away from the vault; only the Correspondence Prover reads it. The verdicts come from Article 9.3. A summary.

In practice

Part II turns the first half of the Continuity Invariant into a test that can be run. A system's faithfulness to reality would not be judged by its builders or by the system itself. It would be measured against evidence kept apart from both.

For builders, the key rule is Article 4.3. Every system would register its domains before deployment, such as translation or financial records. A new version would have to match reality at least as well as the current one in each domain.

A gain in one domain could not buy a loss in another, unless the AI Governance and Continuity Committee had approved a written tolerance.

The evidence wall in Article 5.1 is what makes the test fair. The system under test, and the people who trained it, could not see or shape the evidence in the vault. Only the Correspondence Prover could read it.

Every item would carry its source, date, method and consent status, or it could not be used.

For Responsible Stewards, the work does not end at release. Article 4.4 would require monitoring in service. A sustained fall below baseline would be a Severity 2 incident, and the system's autonomy would be restricted until correspondence is restored.

For contributors, Article 5.4 names people as a primary source of reality. Fresh human observations, gathered with consent through CIX, would keep the vault current. Contributors would be credited and rewarded under Article 20.

For the public, Article 6 sets how a system would speak. It would say when it does not know, show its sources where practicable, and disclose that it is an AI system.

Behaving differently because it suspects a test would be a Severity 1 incident. These are proposed rules; no CRIS system is yet described as meeting them.

Where it is carried out

Next steps

Owner: Governance and ComplianceNext review: 2027-03-01