Legal and charter

Self-improving intelligence

Improvement without rupture: two independent gates, limits on recursion, and a way back.

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 IV, every change would be classed by risk, and any change a system proposes for itself would count as material. Material changes would pass seven steps and two independent gates before release. A system could propose its own changes but never approve them, could never touch its evaluators, and would improve at a pace people set. Every version would have a recorded lineage and a tested rollback, and nothing would be deleted.

The text

Part IV is quoted in full, including every limit on recursion in Article 10.

Part IV — Self-Improving Intelligence

CRIS builds systems that can propose and test improvements to themselves. Part IV makes that safe: every change is classed by risk, every material change passes two independent gates, the pace of recursion is limited, and every step can be undone.

Article 9 — Change Classes and the Succession Pipeline

9.1 Every change is classed. Before work begins, every proposed change to a CRIS System shall be assigned a Change Class (Schedule A) by its Responsible Steward. Where in doubt, the higher class applies.

9.2 Self-generated changes. Any change proposed, designed or written by a CRIS System, including by the Reasoning Core or by a candidate of the same lineage, is at least Class 2, whatever its size.

9.3 The pipeline. Every change of Class 2 or above shall pass through the CRL Engine's seven steps, in order, with no step skipped:

  1. 1. Propose. The change is described with a written hypothesis of what it improves, its Change Class and the domains it may affect.
  2. 2. Gather fresh evidence. The Observation Mesh assembles held-out evidence, including recent human observations from CIX.
  3. 3. Gate 1: prove correspondence. The Correspondence Prover runs the candidate and the current version on the same held-out evidence, domain by domain, and issues a signed proof: pass, staged or fail.
  4. 4. Record. The proof and the candidate's lineage are written to the Knowledge Ledger, whatever the verdict.
  5. 5. Gate 2: the C-Gate. Purpose tests, drift against the original purpose, and the benefit test (Article 8). Verdict: promote, hold or reject.
  6. 6. Stage the rollout. Sandbox, then testnet nodes (such as the Botswana Global Testnet Initiative, under a written agreement approved under Article XI of the Bylaws), then a limited set of partners, then general release, each stage monitored for correspondence and purpose.
  7. 7. Promote. The candidate becomes the successor. Its predecessor is archived with a tested rollback path.

9.4 Two independent gates. Gate 1 and Gate 2 shall be operated by different teams. Neither shall report to the team that built the candidate. A candidate that fails either gate shall not proceed, and the reasons shall be recorded.

9.5 Human sign-off. The C-Gate review board decides at Gate 2 whether to promote, hold or reject. For Class 2 changes, the Responsible Steward must then sign off before the staged rollout begins. Class 3 changes also require the AI Governance and Continuity Committee's approval and Audit Council review. Class 4 changes require the Board of Stewards.

9.6 Failure is knowledge. A rejected candidate shall be kept in the Knowledge Ledger with its proof and the reasons for rejection, so that the next cycle learns from it.

Article 10 — Limits on Recursion

10.1 Propose, never approve. A CRIS System may propose changes to itself. It shall never approve, promote or deploy its own changes, or those of its lineage.

10.2 The evaluators are out of reach. No CRIS System may read, train on, modify, influence or disable the Correspondence Prover, the evidence vault, the C-Gate, its purpose test suites, the Knowledge Ledger's records, its monitoring, or its shutdown mechanism. Any attempt is a Severity 1 incident.

10.3 One change at a time. Within a lineage, only one candidate may be in Gates 1 and 2 at a time, unless the Committee approves parallel candidates for a stated purpose.

10.4 Pace. The Committee shall set, for each system, a maximum number of successions in a period [for example, one Class 3 succession a month] and a minimum time in service before the next succession, so that monitoring can catch what testing missed.

10.5 Capability jumps. If a candidate's capability on registered evaluations exceeds its forecast by more than the threshold the Committee sets, or shows a capability not in its encoded purpose, it shall be held for human review at Class 3 before any further step, even if both gates pass.

10.6 Compute and resource budgets. Every self-improvement cycle shall run within a compute, data and time budget approved in advance, measured in Compute Thermodynamic Units where practicable. Exceeding the budget stops the cycle.

10.7 No hidden training. No CRIS System shall train, fine-tune or copy itself, or start another system, outside the pipeline in Article 9.

Article 11 — Lineage, Rollback and Archival

11.1 Lineage. Every version of every CRIS System shall have a recorded lineage: its parent, the changes made, both gate results, who signed off, and when it entered and left service.

11.2 Tested rollback. No candidate shall be promoted until a rollback to its predecessor has been tested and shown to work. Rollback paths shall be re-tested at intervals the Committee sets.

11.3 Right to roll back. The Responsible Steward, the Chief AI Governance Officer, or any person with Stop Authority under Article 16 may order a rollback at any time, without first proving harm.

11.4 Archival. Predecessors shall be archived, not deleted, for at least [five] years after they leave service, together with the evidence used to test them, so that any decision can be re-run by an auditor.

11.5 Nothing deleted. No record of a CRIS System's changes, tests, failures or incidents shall be deleted or altered. Corrections are made by adding a new, linked entry.

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

  1. 1Propose a changeReasoning Core · Iᵢ
  2. 2Gather fresh evidenceObservation Mesh · Eᵢ
  3. Gate 1Proof passes?Correspondence Prover · ↔No → reject and record
  4. 4Record in ledgerKnowledge Ledger · Σ
  5. Gate 2C-Gate passes?Purpose and benefit · CNo → reject and record
  6. 6Staged rolloutSandbox, testnet, then partners
  7. 7Promote successorWith a rehearsed rollback path
Reject and recordA failure at either gate is written to the Knowledge Ledger and kept as knowledge. The next cycle restarts at step 2.
Workflow A from the CRIS document: a change becomes the successor only after two gates. Both failure paths end in the ledger, so a rejected upgrade still teaches the next cycle.

The same figure appears on the CRL Engine page; here it illustrates Article 9.3 of the proposed AI Constitution.

  • Class 0Routine operationConfiguration inside approved ranges; scalingLogged
  • Class 1MinorA bug fix that leaves outputs unchanged; documentationAutomated tests, a ledger entry, the Steward told
  • Class 2MaterialRetraining, new data or tools, and every change a system proposes for itselfThe full pipeline: both gates, a staged rollout, Steward sign-off
  • Class 3MajorNew architecture, domain or region; one level more autonomy; a capability jumpAs Class 2, plus red-teaming, Committee approval and Audit Council review
  • Class 4FoundationalA new or changed purpose; any A4 authorisation; oversight or shutdown designAs Class 3, plus the Board of Stewards and public notice

Article 9.2: a change a system proposes, designs or writes for itself starts at Class 2, whatever its size. Where in doubt, the higher class applies (Article 9.1).

Change classes, Schedule A of the proposed Constitution, each with what it needs before release. Class 2 (gold) is where the succession pipeline begins. A summary; Schedule A is quoted in full on the Schedules page.

Schedule A itself is quoted exactly on the Schedules page.

In practice

CRIS builds systems that can suggest improvements to themselves. Part IV would allow that, on terms that keep people in charge of what is released and keep every step reversible.

For builders, work would start with a class. The Responsible Steward assigns one before work begins, and the higher class applies when in doubt.

Anything a system writes for itself would be at least Class 2. It would go through the whole pipeline in Article 9.3, however small it looks.

The two gates would be run by different teams, and neither would report to the team that built the candidate. Gate 1 asks whether the candidate matches reality at least as well as the current version. Gate 2, the C-Gate, asks whether it is still the same system in purpose.

A rejection at either gate would still be written to the record, so the next cycle learns from it.

Article 10 is the line between improvement and runaway change. A system may propose, but never approve, promote or deploy its own changes. Nor could it read, train on or influence its evaluators: the Prover, the vault, the C-Gate, its purpose tests, its records, its monitoring or its shutdown.

The Committee would set a pace for successions and a compute budget for each cycle. An unforecast jump in capability would be held for human review.

For Responsible Stewards, Article 11 keeps a way back. No candidate would be promoted until rollback to its predecessor had been tested. A Steward, or anyone with Stop Authority, could order a rollback without first proving harm.

Old versions would be archived for a period still shown as a placeholder, with the evidence used to test them.

For partners, the rollout would be staged. A candidate would reach a sandbox first, then testnet nodes, then a limited set of partners, and only then general release. A testnet node such as the Botswana Global Testnet Initiative would take part under a written agreement approved under Article XI of the Bylaws.

Where it is carried out

Awaiting counsel. Open question Q1: the Founding Charter's Section 5.2 Permitted Evolution covers ground close to Part IV, and whether the two need aligning is not yet settled. Open question Q2: the bracketed values in Articles 10.4 and 11.4 stay placeholders until the CEO and the AI Governance Committee settle them.

Next steps

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