Operator · Intelligence

Reasoning Core

Improvement that starts from verified ground.

Approved by the Board for implementation · 20 November 2026

In one paragraph

The Reasoning Core is where a CRIS system works out how to improve. It reads only what has been verified, including every place where its model and the evidence still disagree, ranks ideas for improvement by how much they should help for what they would cost, and drafts the most promising change with a plan for testing it. Then it hands the change over. The Core never tests, approves or deploys its own changes: other systems and people do that.

Its full name is the Reasoning and Redesign Core. It carries Iᵢ, internal reasoning and computation, the third operator of the Fourth Law: "the system's model, inference and design capacity". It is the part of a self-improving system that asks, given what has been verified, what should change, and writes that answer down as a proposal.

Where it sits

The Core carries the third stage of the loop, Reason. In the CRL Engine's Workflow A it proposes the change at step 1; two gates follow before anything reaches the world.

Knowledge today.
A more continuous tomorrow.
  1. Observe

    Observe the world as it is, and record where each observation came from.

    Learn more about observe →

  2. Verify

    Verify every observation, and every change a system makes to itself, against evidence it has not seen.

    Learn more about verify →

  3. Reason

    Reason only from what has been verified, in steps someone else can trace.

    Learn more about reason →

  4. Accumulate

    Accumulate what was verified, so knowledge adds up instead of being overwritten.

    Learn more about accumulate →

  5. Benefit

    Benefit people and the environment: participants share in the value they create, and every change is weighed against our public benefit purpose.

    Learn more about benefit →

  6. Continue

    Continue, changing anything except correspondence with reality and continuity of purpose.

    Learn more about continue →

  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.

Verified ground

Our rule is simple: no reasoning before verification. The Core starts only from knowledge that has passed the Correspondence Prover and from the contradictions the Prover has found. It does not learn from its own outputs, and it never sees the evidence used to test it.

Reads

Verified knowledgeKnowledge Ledger entries promoted to “verified”
Open contradictionsListed by the Correspondence Prover
Correspondence stateThe latest signed proof
Earlier attemptsLedger lineage, including failures

Never reads

Its own outputsLearning from itself leads to self-referential collapse (Chapter 7.1)
Held-out evidenceThe vault used to test it stays closed to it
Unverified claimsNothing that has not passed the Prover
No reasoning before verification: redesign starts only after correspondence holds (book Chapter 9.1).
InputWhere it comes fromWhy
Verified knowledgeKnowledge Ledger, entries promoted to "verified"Reason only from what has passed
Open contradictionsCorrespondence Prover, via the ledgerThe most valuable things to fix
Correspondence stateThe latest signed proofConfirms the ground is verified
Earlier attemptsLedger lineage, including failuresAvoids repeating what failed

Open contradictions are the Core's main source of work: each place where the model and the world disagree is a hypothesis waiting to be tested. Reasoning from the system's own outputs instead would lead to what the book calls self-referential collapse (Chapter 7.1).

How it works

Every improvement starts with the same four steps.

  1. 1ReadLatest verified knowledge and open contradictions
  2. 2RankHypotheses by expected gain per CTU
  3. 3DraftThe change and its test plan
  4. 4SubmitTo Workflow A; the Core never deploys
TriggerA verified correspondence state
OutputA proposal, not a new version
NeverTests, approves or deploys its own changes
Every improvement starts the same way (CRIS document, Section 9).
  1. Read the ledger's latest verified knowledge and open contradictions.
  2. Rank hypotheses for improvement by expected correspondence gain per CTU of compute.
  3. Draft the change and its test plan.
  4. Submit it to Workflow A. The Core never deploys its own changes.

The trigger is a verified correspondence state. The output is a proposal, not a new version: a new version exists only after the CRL Engine promotes it.

Ranking by gain per CTU

Every idea is weighed by the improvement it should bring against the compute and energy it would use, measured in Compute Thermodynamic Units (CTU). Fixing a known error cheaply comes before an expensive change with an uncertain gain. The chart is an illustration, not a real result.

Illustrative — not real hypotheses or results
  1. 1Fix a known contradiction in one domainSmallLowProposed first
  2. 2Retrain on newly verified evidenceMediumMediumProposed next
  3. 3Redesign a whole moduleUncertainHighRanked lower
  4. 4An idea with no verified basis——Not proposed
Each idea is weighed by how much closer to reality it should bring the system, divided by its cost in Compute Thermodynamic Units. Fixing a known error cheaply comes before an expensive change with an uncertain gain.

This keeps improvement honest about its price. A large change that costs a great deal for a small gain ranks below a small fix to a known contradiction, and an idea with no verified basis is not proposed at all.

The proposal

What the Core hands over is a complete package, so the systems that check it know exactly what is claimed and how to test it.

Proposal package
  1. 1The changeTo methods, architecture, training data or parameters
  2. 2Written hypothesisWhat it should improve, and in which domains
  3. 3Test planThe domains and evidence types the Prover should use
  4. 4BasisThe ledger entries and contradictions it starts from
  5. 5Expected costCompute and energy, in CTU
  6. 6LineageParent version and earlier attempts, including failures
What the Core submits for checking: a proposal, not a new version. A new version exists only after the CRL Engine promotes it.
PartWhat it contains
The changeTo methods, architecture, training data or parameters
Written hypothesisWhat it should improve, and in which domains
Test planThe domains and evidence types the Prover should use
BasisThe ledger entries and contradictions it starts from
Expected costCompute and energy, in CTU
LineageParent version and earlier attempts, including failures

Limits

The Core may propose changes to how a system reasons, how it is built, what it learns from and its settings. It may never change the system's purpose, the gates that check it, its recorded history, or its own access to test evidence. Those limits are enforced by the systems around it.

May propose changes to

  • How the system reasons (methods)
  • How it is built (architecture)
  • What it learns from (training data)
  • Its settings (parameters)

May never change

  • The system’s purpose
  • The gates that check it
  • The ledger’s recorded history
  • Its own access to test evidence
What the Core may and may never change. The limits are held by the systems around the Core, not by its own judgment, as the Continuity Invariant requires (book Chapter 6).

Separation of duties

No single system proposes, tests and approves the same change. The Core proposes; the Observation Mesh gathers fresh evidence; the Correspondence Prover and the C-Gate decide; the CRL Engine rolls the change out in stages; and the Knowledge Ledger records every step.

  1. ProposeReasoning CoreContinuity Labs sets research priorities
  2. Gather evidenceObservation MeshSource owners give consent
  3. Test · Gate 1Correspondence ProverAudit Suite re-runs a sample
  4. Check purpose · Gate 2C-GateReview board for each release
  5. PromoteCRL EngineAudit Council for major versions
  6. RecordKnowledge LedgerAuditors read it
The Core proposes; others decide. No single system proposes, tests and approves the same change.
RoleSystemPeople
ProposeReasoning CoreContinuity Labs sets research priorities
Gather evidenceObservation MeshSource owners give consent
Test (Gate 1)Correspondence ProverAudit Suite re-runs a sample
Check purpose (Gate 2)C-GateReview board for each release
PromoteCRL EngineAudit Council for major versions
RecordKnowledge LedgerAuditors read it

This separation is what lets a self-improving system keep its guards against all four ruptures, rather than marking its own work.

Rupture (Chapter 2)Guard in the loopCRIS system
Correspondence driftThe system's model slowly separates from the world.↔ Correspondence proofEach change tested against realityCorrespondence ProverObservation Mesh
Goal discontinuityIts purpose is reinterpreted at each rewrite until it survives in name only.C Purpose testA successor must keep its purposeC-Gate
Knowledge erosionVerified knowledge is overwritten by unverified assumptions.Σ Promotion rulesOnly verified claims persistKnowledge Ledger
Human severanceThe improvement loop closes around itself and people become bystanders.Eᵢ Human observationPeople stay a source of realityCIXObservation Mesh
The four ways a self-improving system can lose continuity, and the CRIS systems designed to guard against each. Select a guard to see how it is designed to work.

Learning from failure

When a proposal fails, it is recorded with the reasons, and the Core reads that record at the start of its next cycle. A rejected change is not wasted: it tells the next attempt what not to do.

  1. Proposal failsAt Gate 1 or Gate 2
  2. Recorded with reasonsIn the Knowledge Ledger, not deleted
  3. Read next cycleThe Core sees what failed and why
  4. Better proposalThe same change is not resubmitted unchanged
Rejected upgrades are knowledge (book Chapter 8.6): every failure teaches the next attempt what not to do.

Status and read more

This status is set by the CRIS team and links to the evidence behind it. The CRIS roadmap places the first CRL prototype, and with it the Core's first proposals, in 2027.

Current status: Status to be confirmed

LevelEvidence requiredWho sets it
ResearchConcept page and draft proposal format publishedCEO
PrototypeProposals generated from a test ledger and submitted through Workflow A in a test CRL cycleCEO, on the Labs and Systems leads’ advice
PilotProposals used in a partner deployment on testnet; Audit Council review of a sampleCEO
LiveProduction use; external assessment passedCEO, with notice to the board

For developers: Concepts: the loop as data (the proposal format is planned).

Next steps

Owner: Head of Systems EngineeringNext review: 2027-03-01