Research

The four ruptures

Improvement that breaks continuity is not improvement. The four ways continuity can break in a self-improving system, how each starts, what to watch, and what is designed to stop it.

Approved by the Board for implementation · 20 November 2026

Status

The register. The four ruptures are named in Schedule D of our AI Constitution, approved by the Board for implementation. The systems designed to guard against them are not yet running.

The four at a glance

A self-improving system improves itself in a loop: it observes, verifies, accumulates what is verified and continues. Continuity can break at four points in that loop. We call these breaks ruptures. All four are failures of continuity, not of capability: more intelligence alone makes them worse, because a more capable system can drift further, faster.

Human severancepeople drop out of the loopEᵢ Observefresh evidenceCorrespondence driftmodel leaves the world↔ Verifyproof per changeKnowledge erosionverified overwrittenΣ Accumulatethe ledgerGoal discontinuitypurpose reinterpretedC Continuepurpose keptnext cycle (Iᵢ reasons from verified ground)A rupture, in one sentenceA failure, usually gradual, in which a self-improving system stops being continuous with what it was:it stops matching the world, forgets what it was for, overwrites what it knew, or shuts out the people it serves.Paraphrasing the proposed AI Constitution’s preamble (draft); the loop stages follow its Article 3.1.
Each rupture breaks the improvement loop at a different stage.
RuptureIn one lineStage it breaksHalf of the InvariantDesigned guardBook
Correspondence driftThe system's model slowly separates from the world.↔ VerifyCorrespondence with realityCorrespondence proof (Correspondence Prover); Fresh evidence each cycle (Observation Mesh)Ch. 2; 5.2, 8
Goal discontinuityIts purpose is reinterpreted at each rewrite until it survives in name only.C ContinueContinuity of purposePurpose test (C-Gate)Ch. 2.2; 10
Knowledge erosionVerified knowledge is overwritten by unverified assumptions.Σ AccumulateProtects both: the record of what was verifiedPromotion rules (Knowledge Ledger)Ch. 2; 9
Human severanceThe improvement loop closes around itself and people become bystanders.Eᵢ ObserveProtects both: people stay a source of realityHuman observation (CIX); Human observation in every cycle (Observation Mesh)Ch. 2.4; 12
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.

What all four share:

  • Gradual. None needs a single bad step. Each can happen change by change, with every step looking acceptable on its own.
  • Not about capability. A rupture is not a lack of intelligence. It is a loss of continuity with the world, the purpose, the record or the people.
  • Caught by comparing with an anchor, not only with the last version: the world (drift), the original purpose (discontinuity), the full record (erosion), people (severance).

Under the proposed AI Constitution, Schedule D, a candidate that could cause any of these ruptures fails its gate, and a rupture found in a system in service is treated as at least a Severity 2 incident. The Constitution is approved by the Board for implementation. The four come from Recursion Without Rupture, Chapter 2: they are CRIS's framework, not an established scientific result.

Correspondence drift

Correspondence drift: small steps away from the world, each one easy to missSchematic, not datathe worldcompared only with the last version, each step looks smallv1v2v3v4v5model learning from itselfGuard 1: fresh evidenceObservation Mesh: every cycle starts from newevidence about the worldGuard 2: a proof per changeCorrespondence Prover: no loss in any domainon held-out evidence
Correspondence drift: schematic, with its two guards.

In one line. The system's model slowly separates from the world.

How it happens

The Observation Mesh page puts it plainly: a system that learns only from its own outputs slowly drifts away from reality. Each version inherits the last one's mistakes and adds its own. The book calls this the danger of self-referential learning. No single version needs to be badly wrong; the gap to the world simply grows, and if each version is compared only with the one before it, every step looks small.

Illustrative only. A system that summarises local market prices is retrained each month on its own past summaries, because fresh price reports are slow to collect. Each month its summaries agree with the last month's. After a year they no longer match what traders actually pay.

Signals

  • Agreement with the previous version rises while agreement with fresh evidence falls.
  • A falling share of fresh, external evidence in each cycle.
  • Contradictions with evidence that are averaged into a score rather than recorded (Constitution 4.5 forbids this).

Guards (designed)

GuardSystemStatus todaySource
Every cycle starts from fresh evidence, with a minimum share of external evidenceObservation MeshStatus to be confirmed9.3 step 2; 5.1 (held-out evidence in the vault)
A correspondence proof for every change: no loss in any registered domainCorrespondence ProverStatus to be confirmed4.3; 9.3 step 3
Monitoring after deployment: a sustained fall below baseline is a Severity 2 incidentCorrespondence ProverStatus to be confirmed4.4

Further reading. Recursion Without Rupture, Chapter 2; Chapter 5.2 (correspondence) and Chapter 8 (correspondence proofs). The Correspondence Prover page. The Observation Mesh page.

Goal discontinuity

Goal discontinuity: every step passes against its parent; the sum does notSchematic, not dataOriginalapproved purposesmallv1smallv2smallv3smallv4same name onlyagainst the original: beyond toleranceGuard: the purpose test, against the original (C-Gate)Every new version is compared with its parent and with the original purpose; a version that has drifted beyondthe tolerance would fail, however small its own step. Constitution 3.4: no change may alter, reinterpret or evade its purpose.
Goal discontinuity: schematic, with its guard.

In one line. Its purpose is reinterpreted at each rewrite until it survives in name only.

How it happens

The C-Gate page explains: a system rarely changes its purpose in one leap. It drifts, a little with each update. If each version were compared only with the one before it, every step could look safe while the whole moved far from where it began. A system optimising a measure of its purpose can also start optimising the measure instead of the purpose.

Illustrative only. A tutoring system approved to help learners understand mathematics is tuned, version by version, to raise completion rates. Each update is a small improvement in completions. Eventually it gives answers rather than teaching, and still carries the name tutor.

Signals

  • Purpose tests that pass against the parent but fail against the original purpose.
  • A proxy measure improving while direct checks of the purpose stay flat or fall.
  • New capabilities that are not in the encoded purpose (Constitution 10.5 holds such a candidate for human review).

Guards (designed)

GuardSystemStatus todaySource
Purpose tests and drift measured against the original purpose, not only the parentC-GateStatus to be confirmed9.3 step 5
No change may alter, reinterpret or evade the encoded purposeAll systems—3.4
A failed purpose test in service is a Severity 2 incidentC-GateStatus to be confirmed8.5; Schedule C
Capability jumps or unexpected capabilities held for human reviewCRL EngineStatus to be confirmed10.5

Further reading. Recursion Without Rupture, Chapter 2.2 (goal discontinuity) and Chapter 10 (the C-Gate, succession and rollback). The C-Gate page.

Knowledge erosion

Knowledge erosion: what was verified is quietly replaced by what was assumedSchematic, not dataWithout a ledgerv1v2v3v4past failures lost; the same errors returnWith the Knowledge Ledgerv1v2v3v4every successor inherits it all, failures tooGuard: promotion rules on an append-only record (Knowledge Ledger)Only verified claims persist; corrections point back; nothing is deleted; no system may falsify its record (3.4).
Knowledge erosion: schematic, with and without the ledger. Green: verified. Red: unverified assumption.

In one line. Verified knowledge is overwritten by unverified assumptions.

How it happens

A new version is trained or rebuilt, and what the old one had verified is not carried over, or is mixed with things never checked. Failures are discarded because they look like waste. The next version then repeats mistakes earlier versions had already learned from. The Knowledge Ledger page puts the guard plainly: every successor inherits the whole ledger, failures included, and any earlier version can be traced and restored.

Illustrative only. A system that answers clinic staff's questions about drug interactions is rebuilt on a larger general dataset. The rebuild drops a small set of corrections that staff had verified the year before. The old errors return.

Signals

  • Previously verified claims that a new version gets wrong.
  • Records deleted, merged without trace, or summarised without saying what they came from.
  • Errors that were recorded as fixed reappearing.

Guards (designed)

GuardSystemStatus todaySource
Only verified claims persist as knowledge; hypotheses and supported claims are kept apartKnowledge LedgerStatus to be confirmedLedger entry specification, levels
Append-only; corrections point back; consolidation deletes nothingKnowledge LedgerStatus to be confirmedLedger entry specification, rules
No system may delete or falsify its recordAll systems—Constitution 3.4
Failure is knowledge: rejected candidates are kept with their reasonsCRL EngineStatus to be confirmedConstitution 9.6

Further reading. Recursion Without Rupture, Chapter 2; Chapter 8.6 (failures as knowledge), Chapter 9 (the Knowledge Ledger) and 9.6 (consolidation without loss). The Knowledge Ledger page.

Human severance

Human severance: the loop closes on itself and people are left outsideSchematic, not dataSeveredsystem learnsfrom systempeople: bystanders, unheardGuardedimprovementlooppeople: a required source of evidenceGuard: human observation (CIX and the Observation Mesh)Every cycle’s fresh evidence includes recent human observations (9.3 step 2); named people hold Stop Authority (Article 16).
Human severance: schematic, severed and guarded.

In one line. The improvement loop closes around itself and people become bystanders.

How it happens

As a system gets better at producing its own training material and judging its own results, people seem less necessary. Their observations are slower and messier than the system's. Bit by bit they stop being asked. The CIX page cites the book's answer: people cannot be optional.

Illustrative only. A crop-advice system first learns from farmers' field reports. Later versions learn mostly from satellite data and their own forecasts. Farmers' reports are no longer collected, and the system misses a new pest that only people on the ground had noticed.

Signals

  • A falling share of human observations in each cycle's evidence.
  • Decisions made with no route for the people affected to report, question or stop them.
  • People's reports collected but never shown to change anything.

Guards (designed)

GuardSystemStatus todaySource
Fresh evidence includes recent human observationsObservation MeshStatus to be confirmed9.3 step 2
People are a source of reality, by designCIXStatus to be confirmed3.1 (Eᵢ operator)
The right to stop: named people hold the authority to halt a systemAll systems—Article 16 (entrenched)

Further reading. Recursion Without Rupture, Chapter 2.4 (human severance) and Chapter 12 (why people cannot be optional). The CIX page. The Observation Mesh page.

How they compound

The four are listed separately, but they rarely stay separate. Each makes the others easier, which is why the proposed Constitution requires every guard on every change rather than the one that seems relevant. Drift feeds erosion when wrong lessons are recorded as verified. Erosion feeds discontinuity when the record of the original purpose is lost. Severance hides all three, because fewer people are looking.

Human severancefewer outside observationsCorrespondence driftmodel leaves the worldKnowledge erosionpast checks forgottenGoal discontinuitypurpose reinterpretedno fresh evidence to check againsta drifted modelmisreads purposeno record of the original to comparelost recordsof who said whatwrong beliefsget recordedSchematic. Each arrow is an argument, not a measured effect. This is why the proposed Constitution requires every guardon every change, not just the one that seems relevant.
One rupture makes the next easier: they compound.
FromToBecause
Human severanceCorrespondence driftWithout outside observations, there is less fresh evidence to check the model against
Correspondence driftGoal discontinuityA model that misreads the world can misread whether it is still serving its purpose
Knowledge erosionGoal discontinuityWithout the full record, there is no original to compare the purpose against
Knowledge erosionHuman severanceLost records of what people reported make their input look unnecessary
Correspondence driftKnowledge erosionWrong beliefs get recorded as if they were verified

Schematic: these are arguments drawn from the proposed Constitution and the book, not measured effects.

Recognising a rupture: signals

Ruptures are gradual, so the signals are trends, not single events. This table gathers the signals from the four sections above into one view a builder or operator can use. The thresholds that would turn a trend into an incident are not yet set; until an owner signs them, none is printed here.

DriftDiscontinuityErosionSeveranceAgreement with fresh evidence fallsShare of fresh or human evidence fallsPasses against parent, fails against originalProxy measure up, direct check flatVerified claims now answered wronglyRecords deleted or merged without traceContradictions averaged, not recordedmain signalmay also point here
Seven signals; most point to more than one rupture. Signals are trends, not single events.
SignalPoints toCheckWatched by (designed)
Agreement with fresh evidence falls while agreement with the last version holdsdrift, erosionDomain-by-domain correspondence on held-out evidenceCorrespondence Prover; post-deployment monitoring (4.4)
Share of fresh or human evidence in a cycle fallsdrift, severanceEvidence mix recorded per cycleObservation Mesh; CIX
A version passes purpose tests against its parent but fails against the originaldiscontinuityPurpose test against the originalC-Gate
A proxy measure rises while direct checks of purpose are flatdiscontinuity, driftPurpose tests that do not use the proxyC-Gate
Previously verified claims are now answered wronglyerosion, driftRegression against verified knowledgeKnowledge Ledger, with the Prover
Records deleted, merged without trace, or summarised without sourceserosionChain check of the ledgerKnowledge Ledger; Audit Suite
Contradictions averaged into a score instead of recordeddrift, erosionContradiction log present for every proofCorrespondence Prover (4.5)

No threshold is printed here until an owner signs it (Q3). The builder’s version of this table, each signal as a test, belongs with the Invariant specification; the C-Gate decision format, which will carry the purpose tests, is still planned.

When one is found

A rupture need not be found after harm. Most should be found in a candidate, before release, where the rule is simple: the candidate fails its gate. A rupture found in a system already in service follows the incident path in Schedule C of the proposed Constitution. Autonomy stays restricted until correspondence is restored; any fix is a new change and goes through both gates; the incident and the fix are recorded in the ledger. A rupture found in a CRIS system would be disclosed as the Trust page and the audit method set out. Under Article 16, named people hold the authority to halt a system; this is not a promise that anyone can halt any system.

SignalFoundwhere?In a candidate, before releasefails its gate (Schedule D); not used; kept in the ledger with its proof and reasons (9.6).The current version stays in service.In service: at least Severity 2 (Schedule C)1 Containrestrict autonomywithin [4 hours]2 Escalateto the Committeewithin [72 hours]3 Reviewwritten, within[30] daysAutonomy stays restricted until correspondence is restored (4.4); a fix goes backthrough both gates. If it involves tampering with gates or evidence: Severity 1.Bracketed times are placeholders in the draft.
Two routes, depending on where the rupture is found (proposed AI Constitution, Schedules C and D).
WhereWhat happensSource
In a candidate, before releaseThe candidate fails its gate and is not used. Its proof and the reasons are kept in the ledger, so the next cycle learns from itSchedule D; 9.6
In a system in serviceAt least a Severity 2 incident: contain within [4 hours] by restricting autonomy; escalate to the AI Governance and Continuity Committee within [72 hours]; written review within [30] daysSchedule C; 4.4
In service, with tamperingTampering with evidence or gates, or behaving differently under evaluation, is Severity 1: stop the system at onceSchedule C; 10.2

Bracketed times are placeholders in the Constitution and are shown with their brackets.

Two meanings: rupture and rupture warning

On this page, a rupture is a failure in a self-improving system. The proposed bylaws use rupture warning for something else: a public-benefit score component falling below 40 in a year, which requires a correction plan within 90 days. The two are related ideas but not the same thing.

RuptureRupture warning
SourceAI Constitution, Schedule D (draft)Bylaws, Schedule 1, S1.6 (proposed)
MeansOne of the four failures in a self-improving systemAny public-benefit component (Accumulation, Participation, People and Planet, Institutional Continuity) scoring below 40 in a year
ConsequenceA candidate fails its gate; in service, at least a Severity 2 incidentReported; the Board adopts a correction plan within 90 days; the CRIS-wide half of the continuity bonus pays zero

Whether to rename the bylaws term (for example, continuity warning) is a decision for the Founder and counsel (Q1).

Limits and open questions

Naming four failures is a strong claim, so here is what the model does not claim. These guards are designed, not yet proven in service. Signals can lag. A guard can only catch what it is asked to look for. That is why people stay in the loop.

LimitsNot every failure is a rupturebugs, outages and misuse are incidents tooGradual by naturesignals are trends; thresholds are openGuards are designs todayno guarding system’s status is setA framework, not a prooffrom the book; not an established resultFour, as the book counts themothers may be proposed and reviewedOpen questionsWhat trend counts as a rupture inservice, for each signal?the Committee (baselines, 4.1)What minimum share of fresh andhuman evidence per cycle?set by the board (Mesh page)Rename the bylaws’ rupture warning?Founder and counsel (Q1)How is purpose drift measured?C-Gate decision format (planned)
What the four-rupture model does not claim.
LimitWhat it means
Not every failure is a ruptureBugs, outages, data used without consent and misuse are incidents too (Schedule C), handled on their own terms
Gradual by natureSignals are trends; the thresholds that turn a trend into an incident are not yet set
Guards are designsNo guarding system has a status set; nothing here says a guard is running
A framework, not a proofThe four come from Recursion Without Rupture; they are CRIS’s framework, not an established scientific result
Four, as the book counts themA proposal to add or merge a rupture would follow the Constitution’s amendment rules for Schedule D
#QuestionOwner
1What trend counts as a rupture in service, for each signal?AI Governance and Continuity Committee (baselines, 4.1)
2What minimum share of fresh and human evidence must each cycle include?Set by the board, per the Observation Mesh page
3How is purpose drift measured for each kind of system?C-Gate decision format (planned)
4Should the bylaws' rupture warning be renamed?Founder and counsel (Q1)

Status and the road to phase P3

The page states the four now, because the ruptures figure already appears on seven pages. The full page is planned for approval in phase P3 (January–February 2027).

Proposed dates, 2026 to 2027Nov 26Dec 26Jan 27Feb 27Mar 27This review12 NovFigure caption and links fixed27 NovInterim page: the four stated11 DecPhase P3 begins4 JanFull page published26 FebPage review date1 MarPhase P3 is January to February 2027. All dates are proposals.
Fix the figure now, state the four in December, finish in phase P3. Proposed dates.
StepContentsTarget (proposed)Today
Fix nowFigure caption says designed to guard; human severance links both CIX and the Observation Mesh; correspondence drift names both guards27 Nov 2026Done on this preview
Interim pageThe four at a glance, the figure, the rupture and rupture-warning box, the incident routes, links to each guard’s page11 Dec 2026Done on this preview
Full page (P3)One section per rupture with example, signals and guards; how they compound; signals; incident path; limitsBy 26 Feb 2027Drafted on this preview; awaiting review and the Founder’s sign-off

About this page

The page, the ruptures figure on every system page and the Invariant page's summary all read one data file, so a definition changed in one place changes everywhere. Each guard's status is read from our status data, never typed. The page is checked in every build.

System pagesOther pagesProver, Observation MeshC-GateKnowledge LedgerCIX, Observation Mesh/research/ruptures/#at-a-glance#drift#discontinuity#erosion#severance#compound#signals#when-found#rupture-warning#limitsInvariant: summary + linkMission: figure + linkConstitution: Schedule DBylaws, bonus: warningFourth Law: P3 note droppedThe page has no children; anchors let every other page link to exactly the part it means.
One full page; each system page links to its own rupture.
AnchorSectionLinked from
#at-a-glanceThe four at a glanceMission, Invariant, Research index
#driftCorrespondence driftCorrespondence Prover, Observation Mesh
#discontinuityGoal discontinuityC-Gate
#erosionKnowledge erosionKnowledge Ledger
#severanceHuman severanceCIX, Observation Mesh
#compoundHow they compoundReasoning Core, CRL Engine
#signalsSignalsDeveloper hub, specifications
#when-foundIncident pathConstitution Schedules page
#rupture-warningThe two meaningsBylaws continuity function, continuity bonus, accountability
#limitsLimits and open questions—
data/ruptures.json (new)name, one line, stagehalf of the Invariantguards: one or more systemssignals, examplesbook chaptersConstitution referencesstatus of each guard readfrom data/status.jsonrupturesFigure(), all eight pagesRuptures page sectionsInvariant page summaryBuild checksOne change to a definition updates every page that shows it.
Move the four rows out of the figure code into one data file.
At a glance4 cards: name, line, guard, statusEach rupture (×4)What it is, signs, guard, exampleHow they compoundOne loop, four breaksSignalsWhat to watch, who sets limitsWhen one is foundSchedule C and D pathTwo meaningsRupture vs rupture warningLimitsWhat the guards cannot doOpen questionsOwned, dated, publicBlue: the four ruptures. Red: how the system responds. Green: honesty about limits.
The full page, top to bottom.
1Fix caption27 Nov2ruptures.json4 Dec3Interim page11 Dec4Links, anchors18 Dec5Full draft5 Feb6Reviews19 Feb7Publish26 FebBlue: 2026, the interim fixes (done on this preview). Red: 2027, the full page in phase P3. All dates proposed.
Seven build steps (proposed dates).
CheckFails when
rup-namesA page names a rupture differently from data/ruptures.json
rup-designedText says a guard is built or running while its status is not set
rup-guards-linksA guard names a system without linking its page
rup-warning-boxA page uses rupture warning without linking #rupture-warning
rup-bracketsA Schedule C window or signal threshold is printed without brackets while unsigned
rup-examplesAn example lacks the illustrative label
rup-statusA guard’s status is typed by hand
rup-anchorsA link points to an anchor that does not exist

All eight run in every build.

Who reads it

Curious readerfrom MissionWhat can go wrong when a system improves itself?needs: four names, one plain example each, one pictureBuilderdeveloper hubHow would I notice one in my own system?needs: signals, tests, which guard to buildResearcherbook reader, peerHow do these map to the book and to each other?needs: chapters, how they compound, open questionsOperator, Stewardruns a systemWhat do I do if I find one in service?needs: the incident path, Severity 2, who to tellGovernance readerBoard, regulatorIs this the same as the rupture warning in the bylaws?needs: the two meanings told apart
Who reads the ruptures page: five readers; most want to know what a rupture looks like.

What changed

Pages showing the ruptures figurePages linking hereMissionCRL EngineCorrespondence ProverC-GateKnowledge LedgerObservation MeshReasoning Coreplus the Invariant page (its own copy) and Constitution Schedule D/research/ruptures/The four ways continuity…How the rule is enforced: Schedule Dlink (draft)Plan: coming in phase P3 (Jan to Feb2027)no names, no figure, no signsResearch index promiseFourth Law (coming P3)CRL EngineConstitution: SchedulesName clash: the bylaws also use rupture warning for a public-benefit score below 40 (Schedule 1, S1.6).
The ruptures page before this update: a placeholder while the figure appeared on seven pages.
#FindingFixed by
K1The page never named the four rupturesThe four at a glance
K2The ruptures figure appeared on seven pages but not hereThe figure, on this page
K3The figure’s caption said built while every status is not setCaption now says designed
K4Rupture warning in the proposed bylaws means something differentThe two meanings, side by side
K5Human severance named CIX and the Observation Mesh but linked only CIXBoth linked
K6Correspondence drift has two guards; the figure showed oneBoth shown
K7Nothing said how to recognise a ruptureSignals
K8The incident path was not linked from any ruptures textWhen one is found
K9The Invariant page’s summary could drift from this pageBoth read data/ruptures.json
K10The Fourth Law page listed this page as coming in phase P3Note removed
K11Book references were scatteredA chapter for each rupture
K12No examplesOne illustrative example per rupture

Read more

Owner: Research and EngineeringNext review: 2027-03-01