Operator · Correspondence

Correspondence Prover

Every change tested against evidence it has never seen.

Approved by the Board for implementation · 20 November 2026

In one paragraph

The Correspondence Prover is the first of two gates every CRIS system change must pass. It runs the current version and the proposed one side by side on evidence the system has never seen, and checks, area by area, that the new version matches the real world at least as well as the old. The result is a correspondence proof: not a mathematical certainty, but a signed, measurable demonstration that anyone with access can check and re-run.

It carries ↔, correspondence, the second operator of the Fourth Law: "the relation that must hold between what is observed and what is reasoned". In practice it is the system that stands between a proposed change and the world.

Where it sits

The Prover carries the second stage of the loop, Verify. In the CRL Engine's Workflow A it is Gate 1, at step 3, before the C-Gate.

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.

Rules for this system are proposed in the AI Constitution, Articles 4 and 5: the correspondence requirement and evidence. They are approved by the Board for implementation.

What a proof is

A proof takes the current version of a system, the candidate that would replace it, and evidence the system has never seen, and shows, domain by domain, whether the candidate matches reality at least as well. It is a record that can be checked, not a claim to be trusted.

Inputs

Current versionThe live model
CandidateThe proposed change
Held-out evidenceFrom the mesh’s vault, by domain
TolerancesSet by the board, per domain
Correspondence ProverRuns both versions on the same evidence; compares domain by domain; lists every contradiction

Output · proof object

Candidate
The proposed version
Baseline
The live version
Evidence set
Which held-out evidence was used
Domains
The result in each domain
Contradictions
Every disagreement, with sources
Verdict
Pass, staged or fail
Conditions
Limits and monitoring, if staged
Signed
The Prover’s key and the time
A correspondence proof is a signed record, not a mathematical theorem: “a measurable demonstration that a new version models the world at least as faithfully as the old” (book Chapter 8.1). It is written to the Knowledge Ledger whatever its verdict.
Part of the proofWhat it records
Candidate and baselineWhich version is proposed and which is live
Evidence setWhich held-out evidence from the vault was used
DomainsThe result in each domain the change could affect
ContradictionsEvery place where model and evidence disagree, with sources
VerdictPass, staged or fail
ConditionsFor a staged verdict, the limits and monitoring required
SignatureThe Prover's key and time, so tampering is detectable

Whatever the verdict, the proof is written to the Knowledge Ledger. A failed proof is not wasted: the book calls rejected upgrades "knowledge", because they teach the next cycle what not to do.

How it works

Every proof follows five steps. Both versions face exactly the same evidence, so the comparison is fair.

  1. 1SelectThe domains the change could affect
  2. 2Run bothCandidate and current on the same held-out evidence
  3. 3CompareDomain by domain; a loss fails
  4. 4DetectList every contradiction
  5. 5SignIssue the proof to the ledger
TriggerA candidate change from the CRL Engine, or a contribution from CIX
OutputA signed proof: pass, fail or staged, with conditions
NeverAverage a loss in one domain against a gain in another
Every proof follows the same five steps (CRIS document, Section 9).
  1. Select the evaluation domains the change could affect.
  2. Run both the candidate and the current version on the same held-out evidence.
  3. Compare correspondence domain by domain; a loss in any domain fails the proof unless a board-approved tolerance allows a staged trial.
  4. Detect contradictions and list every conflict between model output and evidence.
  5. Sign the proof and issue it to the Knowledge Ledger.

Held-out evidence

The Prover is the only system allowed to read the held-out vault kept by the Observation Mesh. The system under test never sees it. This is what makes the promise on this page true: every change is tested against evidence it has never seen.

QuarantineNew, verified for format and source
Training poolMay be used by the Reasoning Core to propose and train changes
Held-out vaultRead only by the Correspondence Prover to test candidates; never shown to the system under test
Fair testEach candidate is judged on evidence it has never seen
Held-out evidence lives in a vault the system can never train on. The split is fixed for each cycle and recorded in the Knowledge Ledger, so auditors can confirm no vault item reached training.

Domain by domain

A new version cannot buy a loss in one area with gains in others. If it is less faithful anywhere, it does not pass, however much it improves elsewhere. The chart below is an illustration of the rule, not a real result.

Illustrative — not real results
Water availabilitypass
Crop yieldpass
Disease reportspass
Energy demandpass
Market pricesfail
Transportpass
Verdict: failThe candidate is less faithful on market prices, so it is not promoted, even though five domains improved. The failure is recorded and goes back to the Reasoning Core.
A gain in most domains cannot buy a loss in one.

This rule protects against the slow, hidden drift the book calls correspondence drift, the first of the four ruptures. The Prover is its guard.

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.

Read about correspondence drift on the ruptures page.

Contradictions

When a system's model and the evidence disagree, the Prover does not average the disagreement away. It lists each contradiction with its sources, so the system can learn from it and the people responsible can see it.

Illustrative example
Model says“River level normal”
Evidence saysA sensor and three CIX reports: “flooding”
ContradictionListed in the proof, with sources
Reasoning CoreA hypothesis to fix
Observation MeshTrust scores updated
Knowledge LedgerRecorded
Contradictions are surfaced, never averaged away: “contradiction as signal, not failure” (book Chapter 4.2).

Chapter 4.2 of the book calls contradiction "a signal, not a failure": a disagreement between model and world is how a system learns. The Prover surfaces each one, the Reasoning Core treats it as a hypothesis to fix, the Observation Mesh adjusts the trust of the sources involved, and the Knowledge Ledger keeps the record.

Verdicts

Every proof ends in one of three verdicts. Pass: the change goes on to the C-Gate. Staged: a small loss within a margin our board has approved allows a limited, monitored trial. Fail: the change is not used, and the failure is recorded so the next attempt can learn from it.

PassAt least as faithful in every domain

The candidate goes on to the C-Gate (Gate 2).

StagedWithin a board-approved tolerance

A limited, monitored trial with conditions; proved again before promotion.

FailLess faithful in any domain

Not promoted; recorded as knowledge; sent back with the contradictions.

Three verdicts, each with a defined next step. Uncertain domains use probabilistic proofs with a stated confidence (book Chapter 8.4). Tolerances are set by the board, never by the Prover.
VerdictConditionNext step
PassAt least as faithful in every domainCandidate goes to the C-Gate
StagedA small loss within a tolerance the board has approvedLimited, monitored trial; proved again before promotion
FailLess faithful in any domain, beyond toleranceNot promoted; recorded; sent back with the contradictions

For domains where evidence is uncertain, the Prover issues probabilistic proofs with a stated confidence (book Chapter 8.4). Tolerances and confidence thresholds are approved by the board, as part of governing the loop rather than each version. The Prover never sets them.

Two jobs and who checks it

The same method checks contributions on the Continuity Intelligence Exchange: each observation is compared with other evidence before it is accepted. And the Prover is checked in turn: every proof is signed, a sample is re-run independently by the Continuity Audit Suite, people review the proofs behind each release, and the method is assessed against an external standard.

The same method serves the CRL Engine and CIX.
  1. 1Signed proofsEvery proof is signed and stored, so tampering is detectable
  2. 2Independent re-runsThe Continuity Audit Suite re-runs a sample of proofs
  3. 3Human reviewThe C-Gate review board and the Audit Council read major proofs
  4. 4Third-party standardExternal assessment of the method
Who verifies the verifier? Four independent checks on the Prover itself (book Chapter 8.5).
CheckWhoHow often
Signed proofsThe Prover, automaticallyEvery proof
Independent re-runsContinuity Audit SuiteA sample of proofs, on schedule
Human reviewC-Gate review board; Audit Council for major versionsEvery release; major versions
Third-party standardExternal assessorYearly, with the benefit report

The Prover is the first of the five layers of assurance around every CRIS system.

  1. Automated checksProofs and drift tests
  2. Human reviewC-Gate board, every release
  3. Independent auditAudit Council, major versions
  4. External standardThird-party certification
  5. Public reportBenefit report, every year
Five layers of assurance around the engine, from machine speed to public view.

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 first real proofs, in 2027.

Current status: Status to be confirmed

LevelEvidence requiredWho sets it
ResearchConcept page and draft proof format publishedCEO
PrototypeProofs issued for a test CRL cycle using held-out evidence; proofs stored in the ledgerCEO, on the Labs and Systems leads’ advice
PilotProofs used in a partner deployment on testnet; Audit Suite re-runs a sample; Audit Council reviewCEO
LiveProduction use; external assessment of the method passedCEO, with notice to the board

For developers: the correspondence proof format, draft 0.1, and the Verify path.

Next steps

Owner: Head of Continuity LabsNext review: 2027-03-01