Continuity Audit Suite
Tools for anyone who needs to check our work.
In one paragraph
The Continuity Audit Suite is how anyone can check CRIS's work. It samples the proofs that test every change, the decisions that promote or reject new versions, and the permanent record of what was learned. It re-runs a share of those proofs independently, and checks every claim in our public benefit report against its evidence. Its findings go to our board and, once it is appointed, to an independent Audit Council, and a summary is published. It is our answer to the oldest question in governance: who checks the checkers?
It does not improve the loop; it checks that the loop did what it says. Every other promise on this site, that changes are tested, that purpose is kept, that nothing is deleted, that participants are paid fairly, rests on someone outside the team being able to check it.
Where it sits
The Audit Suite does not carry one stage of the loop: it checks the work of every stage, and it runs beside every layer of the CRIS stack.
checks every stage
Observe
Observe the world as it is, and record where each observation came from.
Verify
Verify every observation, and every change a system makes to itself, against evidence it has not seen.
Reason
Reason only from what has been verified, in steps someone else can trace.
Accumulate
Accumulate what was verified, so knowledge adds up instead of being overwritten.
Benefit
Benefit people and the environment: participants share in the value they create, and every change is weighed against our public benefit purpose.
Continue
Continue, changing anything except correspondence with reality and continuity of purpose.
Select a tile to see what it does. Every tile links to its page.
Our approach treats safety as continuity: the risk is drift, not intelligence (book Chapter 16.2). Every safeguard checks whether correspondence with reality and purpose survived a change, and the Audit Suite is the layer that checks all the other safeguards. It sits at the independent end of the five layers of assurance.
- Automated checksProofs and drift tests
- Human reviewC-Gate board, every release
- Independent auditAudit Council, major versions
- External standardThird-party certification
- Public reportBenefit report, every year
Rules for this system are proposed in the AI Constitution, Article 22: audit and assurance. They are approved by the Board for implementation.
What it checks
The Audit Suite checks four things, each produced by a different part of CRIS.
- Correspondence proofsFrom: Correspondence ProverDoes the proof hold if re-run independently?
- C-Gate decisionsFrom: C-GateWere purpose tests, drift checks and rollback done as the rules require?
- Knowledge LedgerFrom: Knowledge LedgerIs history complete and unaltered?
- Benefit claimsFrom: Public benefit reportIs every published result backed by evidence?
| What is checked | Produced by | The question | Evidence |
|---|---|---|---|
| Correspondence proofs | Correspondence Prover | Does the proof hold if re-run independently? | Held-out evidence; signed proofs |
| C-Gate decisions | C-Gate | Were purpose tests, drift checks and rollback done as the rules require? | C-Gate reports; review board minutes |
| The ledger | Knowledge Ledger | Is history complete and unaltered? | Entry links; signatures; consolidation records |
| Benefit claims | Public benefit report | Is every published result backed by evidence? | Scorecard sources; payout data; programme results |
How it works
Audits happen on a published calendar, before or after every major version, and whenever someone complains. Each audit follows four steps.
Three triggers
- Scheduled auditOn a published calendar
- Major versionBefore or after release
- ComplaintFrom anyone, inside or outside CRIS
- 1SampleProofs, gate decisions and ledger entries
- 2Re-runA subset of proofs, independently
- 3CheckBenefit claims against their evidence
- 4ReportFindings to the board; a summary published
- Sample proofs, gate decisions and ledger entries.
- Re-run a subset of proofs independently.
- Check benefit claims against their evidence.
- Report findings to the board and publish a summary.
The CRIS plan proposes Audit Council summaries twice a year and an audited benefit report every year.
Sampling and re-runs
The auditor, not the team being audited, chooses what to check: every high-risk item, and a random share of everything else. High-risk items are always included: major versions, decisions held for people, complaints and large payouts. The rest are drawn at random, and the sample is recorded before work starts, so management cannot change it.
- Major versions
- Decisions held for people
- Complaints
- Large payouts
The strongest check is a re-run. The auditor takes the same inputs, the candidate, the baseline, the held-out evidence and the tolerances, and runs the published procedure again on separate systems. A match within the stated tolerance confirms the proof; a mismatch is a finding, and the promotion it supported is reviewed.
- The candidate
- The baseline
- The held-out evidence set
- The tolerances
- Match, within toleranceThe proof is confirmed
- MismatchA finding; the promotion it supported is reviewed
Benefit claims and standards
As a public benefit corporation, CRIS reports on the public benefit it creates. Our charter ties that benefit to a third-party standard, and we publish a report every year. The Audit Suite traces every number in that report to its source and recomputes it, because our risk register names the danger: benefit claims overstated (book Chapter 13.4).
- Published resultA number in the public benefit report
- Trace to sourceScorecard sources, payout data, programme results
- RecomputeThe auditor recalculates it
- Confirmed or a findingOverstated claims are corrected
Two standards are under consideration: ISO/IEC 42001 for the management of AI systems, and B Lab's assessment for overall benefit, with the Audit Suite as our own method mapped to both. These are candidates; the board will record its choice in our bylaws. Results are tracked on our scorecard.
Who audits the auditors
A verifier that answers to the people it checks is not independent. Our plan is an independent Audit Council of three to five members, drawn from academics, standards experts and community representatives, that reviews the results and reports to our board, not to management. An external assessor checks us against a recognised standard. The team that runs the audit tools cannot change the findings.
- BoardReceives findings, requires fixes, publishes the reportReports to: Stockholders, under the public benefit duty
- Audit Council (to be appointed)Three to five independent members; inspects proofs, gate decisions and the ledger; reviews major versions and the benefit reportReports to: The board, not management
- External assessorChecks CRIS against the chosen third-party standardReports to: The board
- Trust and SafetyRuns the tools, the calendar and the toolkit; cannot edit findingsReports to: Management, with a direct line to the Audit Council
| Role | Does | Reports to | Cannot |
|---|---|---|---|
| Audit Council (to be appointed) | Inspects proofs, C-Gate decisions and the ledger; reviews major versions and the benefit report | The board | Be managed by the teams it audits |
| Trust and Safety | Runs the tools, the audit calendar and the toolkit | Management, with a direct line to the Audit Council | Edit findings |
| External assessor | Checks CRIS against the chosen third-party standard | The board | Rely only on CRIS's own reports |
| Board | Receives findings, requires fixes and publishes the report | Stockholders, under the public benefit duty | Delegate the decision to publish |
Findings and disclosure
Every finding is rated, fixed and checked again. A critical finding freezes releases until it is fixed and is disclosed on our Trust page. Every finding appears, at least in summary, in the next public benefit report.
- CriticalA promoted version whose proof does not re-runFixPromotions frozen at once; fixed before any further releaseDisclosureIncident note on the Trust page; in the next benefit report
- MajorA C-Gate decision made without a required rollback testFixWithin one release cycleDisclosureIn the next Audit Council summary
- MinorA ledger entry missing a field, with no effect on decisionsFixWithin the next audit periodDisclosureIn the summary, in aggregate
A critical finding follows our incident response: detect, contain, record, investigate, fix the loop, and disclose.
The public toolkit
You will be able to read our audit method, look at summary views of the record, re-run a published proof with our kit, read a sample report, ask for an audit or make a complaint. Requesting an audit and making a complaint are open now; the rest are linked here as they are published.
- Audit methodThe public form of the book’s continuity audit frameworkDraft 0.1
- Ledger viewsSummaries for the public; full read-only views for accredited auditorsTo be published
- Re-run kitRe-run a published proof yourselfTo be published
- Sample audit reportWhat a report looks likeTo be published
- Request an auditAsk for an audit of something we have doneAvailable now
- Make a complaintA complaint can start an auditAvailable now
Status and read more
This status is set by the CRIS team and links to the evidence behind it. The CRIS plan calls for the first public benefit report to be audited against the chosen standard in 2028.
Current status: Status to be confirmed
| Level | Evidence required | Who sets it |
|---|---|---|
| Research | Concept page; draft audit method (book Appendix D) published | CEO |
| Prototype | A test audit of a CRL test cycle: sample drawn, proofs re-run, findings reported | CEO, on the Labs and Systems leads’ advice |
| Pilot | Audit Council appointed; first audit summary published; first benefit report audited | CEO |
| Live | Regular audits on the published calendar; external assessment against the chosen standard passed | CEO, with notice to the board |
For developers: the Verify path and the audit method, draft 0.1.
- Our public benefit report, scorecard and Trust page.
- Recursion Without Rupture: Chapter 8.5 (who verifies the verifier), Chapter 13.4 (benefit claims), Chapter 16.2 (safety as continuity) and Appendix D (the continuity audit framework).
- Related systems: Correspondence Prover, C-Gate, Knowledge Ledger, Participation Value Meter.
