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.
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.
| Rupture | In one line | Stage it breaks | Half of the Invariant | Designed guard | Book |
|---|---|---|---|---|---|
| Correspondence drift | The system's model slowly separates from the world. | ↔ Verify | Correspondence with reality | Correspondence proof (Correspondence Prover); Fresh evidence each cycle (Observation Mesh) | Ch. 2; 5.2, 8 |
| Goal discontinuity | Its purpose is reinterpreted at each rewrite until it survives in name only. | C Continue | Continuity of purpose | Purpose test (C-Gate) | Ch. 2.2; 10 |
| Knowledge erosion | Verified knowledge is overwritten by unverified assumptions. | Σ Accumulate | Protects both: the record of what was verified | Promotion rules (Knowledge Ledger) | Ch. 2; 9 |
| Human severance | The improvement loop closes around itself and people become bystanders. | Eᵢ Observe | Protects both: people stay a source of reality | Human observation (CIX); Human observation in every cycle (Observation Mesh) | Ch. 2.4; 12 |
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
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.
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)
| Guard | System | Status today | Source |
|---|---|---|---|
| Every cycle starts from fresh evidence, with a minimum share of external evidence | Observation Mesh | Status to be confirmed | 9.3 step 2; 5.1 (held-out evidence in the vault) |
| A correspondence proof for every change: no loss in any registered domain | Correspondence Prover | Status to be confirmed | 4.3; 9.3 step 3 |
| Monitoring after deployment: a sustained fall below baseline is a Severity 2 incident | Correspondence Prover | Status to be confirmed | 4.4 |
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
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.
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)
| Guard | System | Status today | Source |
|---|---|---|---|
| Purpose tests and drift measured against the original purpose, not only the parent | C-Gate | Status to be confirmed | 9.3 step 5 |
| No change may alter, reinterpret or evade the encoded purpose | All systems | — | 3.4 |
| A failed purpose test in service is a Severity 2 incident | C-Gate | Status to be confirmed | 8.5; Schedule C |
| Capability jumps or unexpected capabilities held for human review | CRL Engine | Status to be confirmed | 10.5 |
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
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.
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)
| Guard | System | Status today | Source |
|---|---|---|---|
| Only verified claims persist as knowledge; hypotheses and supported claims are kept apart | Knowledge Ledger | Status to be confirmed | Ledger entry specification, levels |
| Append-only; corrections point back; consolidation deletes nothing | Knowledge Ledger | Status to be confirmed | Ledger entry specification, rules |
| No system may delete or falsify its record | All systems | — | Constitution 3.4 |
| Failure is knowledge: rejected candidates are kept with their reasons | CRL Engine | Status to be confirmed | Constitution 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
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.
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)
| Guard | System | Status today | Source |
|---|---|---|---|
| Fresh evidence includes recent human observations | Observation Mesh | Status to be confirmed | 9.3 step 2 |
| People are a source of reality, by design | CIX | Status to be confirmed | 3.1 (Eᵢ operator) |
| The right to stop: named people hold the authority to halt a system | All 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.
| From | To | Because |
|---|---|---|
| Human severance | Correspondence drift | Without outside observations, there is less fresh evidence to check the model against |
| Correspondence drift | Goal discontinuity | A model that misreads the world can misread whether it is still serving its purpose |
| Knowledge erosion | Goal discontinuity | Without the full record, there is no original to compare the purpose against |
| Knowledge erosion | Human severance | Lost records of what people reported make their input look unnecessary |
| Correspondence drift | Knowledge erosion | Wrong 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.
| Signal | Points to | Check | Watched by (designed) |
|---|---|---|---|
| Agreement with fresh evidence falls while agreement with the last version holds | drift, erosion | Domain-by-domain correspondence on held-out evidence | Correspondence Prover; post-deployment monitoring (4.4) |
| Share of fresh or human evidence in a cycle falls | drift, severance | Evidence mix recorded per cycle | Observation Mesh; CIX |
| A version passes purpose tests against its parent but fails against the original | discontinuity | Purpose test against the original | C-Gate |
| A proxy measure rises while direct checks of purpose are flat | discontinuity, drift | Purpose tests that do not use the proxy | C-Gate |
| Previously verified claims are now answered wrongly | erosion, drift | Regression against verified knowledge | Knowledge Ledger, with the Prover |
| Records deleted, merged without trace, or summarised without sources | erosion | Chain check of the ledger | Knowledge Ledger; Audit Suite |
| Contradictions averaged into a score instead of recorded | drift, erosion | Contradiction log present for every proof | Correspondence 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.
| Where | What happens | Source |
|---|---|---|
| In a candidate, before release | The candidate fails its gate and is not used. Its proof and the reasons are kept in the ledger, so the next cycle learns from it | Schedule D; 9.6 |
| In a system in service | At 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] days | Schedule C; 4.4 |
| In service, with tampering | Tampering with evidence or gates, or behaving differently under evaluation, is Severity 1: stop the system at once | Schedule 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.
| Rupture | Rupture warning | |
|---|---|---|
| Source | AI Constitution, Schedule D (draft) | Bylaws, Schedule 1, S1.6 (proposed) |
| Means | One of the four failures in a self-improving system | Any public-benefit component (Accumulation, Participation, People and Planet, Institutional Continuity) scoring below 40 in a year |
| Consequence | A candidate fails its gate; in service, at least a Severity 2 incident | Reported; 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.
| Limit | What it means |
|---|---|
| Not every failure is a rupture | Bugs, outages, data used without consent and misuse are incidents too (Schedule C), handled on their own terms |
| Gradual by nature | Signals are trends; the thresholds that turn a trend into an incident are not yet set |
| Guards are designs | No guarding system has a status set; nothing here says a guard is running |
| A framework, not a proof | The four come from Recursion Without Rupture; they are CRIS’s framework, not an established scientific result |
| Four, as the book counts them | A proposal to add or merge a rupture would follow the Constitution’s amendment rules for Schedule D |
| # | Question | Owner |
|---|---|---|
| 1 | What trend counts as a rupture in service, for each signal? | AI Governance and Continuity Committee (baselines, 4.1) |
| 2 | What minimum share of fresh and human evidence must each cycle include? | Set by the board, per the Observation Mesh page |
| 3 | How is purpose drift measured for each kind of system? | C-Gate decision format (planned) |
| 4 | Should 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).
| Step | Contents | Target (proposed) | Today |
|---|---|---|---|
| Fix now | Figure caption says designed to guard; human severance links both CIX and the Observation Mesh; correspondence drift names both guards | 27 Nov 2026 | Done on this preview |
| Interim page | The four at a glance, the figure, the rupture and rupture-warning box, the incident routes, links to each guard’s page | 11 Dec 2026 | Done on this preview |
| Full page (P3) | One section per rupture with example, signals and guards; how they compound; signals; incident path; limits | By 26 Feb 2027 | Drafted 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.
| Anchor | Section | Linked from |
|---|---|---|
#at-a-glance | The four at a glance | Mission, Invariant, Research index |
#drift | Correspondence drift | Correspondence Prover, Observation Mesh |
#discontinuity | Goal discontinuity | C-Gate |
#erosion | Knowledge erosion | Knowledge Ledger |
#severance | Human severance | CIX, Observation Mesh |
#compound | How they compound | Reasoning Core, CRL Engine |
#signals | Signals | Developer hub, specifications |
#when-found | Incident path | Constitution Schedules page |
#rupture-warning | The two meanings | Bylaws continuity function, continuity bonus, accountability |
#limits | Limits and open questions | — |
| Check | Fails when |
|---|---|
rup-names | A page names a rupture differently from data/ruptures.json |
rup-designed | Text says a guard is built or running while its status is not set |
rup-guards-links | A guard names a system without linking its page |
rup-warning-box | A page uses rupture warning without linking #rupture-warning |
rup-brackets | A Schedule C window or signal threshold is printed without brackets while unsigned |
rup-examples | An example lacks the illustrative label |
rup-status | A guard’s status is typed by hand |
rup-anchors | A link points to an anchor that does not exist |
All eight run in every build.
Who reads it
What changed
| # | Finding | Fixed by |
|---|---|---|
| K1 | The page never named the four ruptures | The four at a glance |
| K2 | The ruptures figure appeared on seven pages but not here | The figure, on this page |
| K3 | The figure’s caption said built while every status is not set | Caption now says designed |
| K4 | Rupture warning in the proposed bylaws means something different | The two meanings, side by side |
| K5 | Human severance named CIX and the Observation Mesh but linked only CIX | Both linked |
| K6 | Correspondence drift has two guards; the figure showed one | Both shown |
| K7 | Nothing said how to recognise a rupture | Signals |
| K8 | The incident path was not linked from any ruptures text | When one is found |
| K9 | The Invariant page’s summary could drift from this page | Both read data/ruptures.json |
| K10 | The Fourth Law page listed this page as coming in phase P3 | Note removed |
| K11 | Book references were scattered | A chapter for each rupture |
| K12 | No examples | One illustrative example per rupture |
Read more
- The Continuity Invariant: the rule the guards protect.
- AI Constitution: Schedules: the Rupture Register and incident severity.
- Recursion Without Rupture, Chapter 2: the four ruptures.
- The guarding systems: Correspondence Prover, Observation Mesh, C-Gate, Knowledge Ledger, CIX.
