A named category, with a written specification

Claim maintenance

The continuous, independent observation of the public claims an organisation makes about itself or its products: capturing each claim verbatim with its source and date, hashing and timestamping it so the record cannot be quietly rewritten, re-reading it on a schedule, recording every observed change without alleging anything, and measuring the claim against public evidence where — and only where — public evidence can settle it.

Tell us which public claim and source you want to follow. A monitoring scope and cadence would be agreed separately; being listed in this public register does not mean a service has been commissioned.

The specification is dedicated to the public domain under CC0 1.0 — adopt it, fork it or translate it without asking us. The reference implementation is MIT. Archived with a persistent identifier we do not control: 10.5281/zenodo.22901908 (all versions). A DOI makes a document citable and permanent; it does not make it right.

Cite as: Council of AI. Claim Maintenance, version 0.1. CSOAI Ltd, 2026-09-22. https://doi.org/10.5281/zenodo.22901908

Subjects

28

Claims

99

Measured

33

Unmeasured

39

What this does not do

Said first, because the category is easy to mistake for four older ones and because the discipline is the product.

  • It is not fact-checking. No verdict is reached about any claim. Where public evidence exists, the claim and the measured value are published side by side; where it does not, that is said plainly.
  • It is not certification. No mark, badge, seal or grade arises from any state. There is nothing here to pass or fail, and nobody may display one.
  • It is not auditing. No engagement, no agreed scope, no access to non-public records, no opinion addressed to anyone.
  • It is not reputation scoring. No score, rank or index may be derived from the state counts. The states describe the evidence available, not the claimant.
  • It is not adversarial journalism. No theory of anyone's intent is held or implied. The subject-selection rule is published, because that is the one place an agenda can hide.
  • It asserts no falsity about anyone. Nothing published here states or implies that a claim is false, misleading or dishonest — and a listing is neither an endorsement nor an accusation.

Four states, and only four

CLAIM_CAPTURED
Read from its public source, recorded verbatim, hashed, dated and scheduled for re-reading. Nothing has been measured.
CLAIM_MEASURED
A measurement against public evidence exists beside the claim, with its window, denominator, method and result. Not a verdict — the claim and the measurement sit side by side and the reader does the subtraction.
UNMEASURED
In scope, settleable in principle by public evidence, and not yet measured. A first-class state, published as prominently as any other.
UNCHECKABLE
No public evidence can settle it. A statement about the evidence available, never about the subject.

The boundary that makes this honest is the one between an observed change and a finding. An observed change is: the bytes at this URL differ from the bytes recorded at the previous read. That is the whole statement. A finding is: a measurement was performed and produced this value, over this window, with this denominator, by this method. A record of the first is never written in the vocabulary of the second, and never in the vocabulary of motive.

The register — every subject we maintain claims on

Generated from the registry files on disk, never hand-listed, and counting what exists rather than what we wish existed. A small register that grows is worth more than a padded one. This page is showing the register as of 2026-09-25.

SubjectClaimsStatesFirst capturedLast readNext readRegistry
Anthropic2UNMEASURED 22026-09-232026-09-252026-09-28artifact
Arize AI3CLAIM_MEASURED 1 · UNMEASURED 22026-09-232026-09-252026-09-28artifact
Armilla2UNMEASURED 1 · UNCHECKABLE 12026-09-232026-09-252026-09-28artifact
Council of AI (CSOAI Ltd)4CLAIM_MEASURED 1 · UNMEASURED 32026-09-232026-09-252026-09-28artifact
Credo AI3CLAIM_MEASURED 2 · UNMEASURED 12026-09-232026-09-252026-09-28artifact
European Central Bank (Pontes)4CLAIM_MEASURED 1 · UNMEASURED 32026-09-232026-09-252026-09-28artifact
Galileo2UNMEASURED 22026-09-232026-09-252026-09-28artifact
Giskard3CLAIM_MEASURED 1 · UNMEASURED 22026-09-232026-09-232026-09-28artifact
Guardrails AI2UNMEASURED 1 · UNCHECKABLE 12026-09-232026-09-232026-09-28artifact
Helicone3CLAIM_MEASURED 2 · UNMEASURED 12026-09-232026-09-252026-09-28artifact
Holistic AI3CLAIM_MEASURED 2 · UNCHECKABLE 12026-09-232026-09-252026-09-28artifact
InstaHuman4CLAIM_MEASURED 1 · UNMEASURED 1 · UNCHECKABLE 22026-09-242026-09-252026-10-01artifact
Invariant Guardrails3CLAIM_MEASURED 1 · UNMEASURED 22026-09-232026-09-232026-09-28artifact
Invoke5CLAIM_MEASURED 1 · UNCHECKABLE 42026-09-242026-09-252026-10-01artifact
Langfuse5CLAIM_MEASURED 3 · UNMEASURED 1 · UNCHECKABLE 12026-09-232026-09-252026-09-28artifact
Mistral AI2UNMEASURED 1 · UNCHECKABLE 12026-09-232026-09-252026-09-28artifact
NVIDIA NeMo Guardrails4UNMEASURED 2 · UNCHECKABLE 22026-09-232026-09-232026-09-28artifact
Payman5CLAIM_MEASURED 1 · UNCHECKABLE 42026-09-242026-09-252026-10-01artifact
Purple Llama3CLAIM_MEASURED 1 · UNMEASURED 22026-09-232026-09-232026-09-28artifact
PyRIT1CLAIM_MEASURED 12026-09-232026-09-232026-09-28artifact
RentAHuman9CLAIM_MEASURED 2 · UNMEASURED 3 · UNCHECKABLE 42026-09-242026-09-252026-10-01artifact
Securities and Exchange Board of India (Demat 2.0)3CLAIM_MEASURED 1 · UNMEASURED 1 · UNCHECKABLE 12026-09-232026-09-232026-09-28artifact
Traceloop3CLAIM_MEASURED 1 · UNMEASURED 22026-09-232026-09-252026-09-28artifact
Vanta4CLAIM_MEASURED 1 · UNMEASURED 2 · UNCHECKABLE 12026-09-232026-09-252026-09-28artifact
chainlink5CLAIM_MEASURED 4 · UNMEASURED 12026-09-222026-09-22UNSCHEDULEDartifact
garak3CLAIM_MEASURED 1 · UNMEASURED 1 · UNCHECKABLE 12026-09-232026-09-232026-09-28artifact
gotoHuman6CLAIM_MEASURED 3 · UNMEASURED 1 · UNCHECKABLE 22026-09-242026-09-252026-10-01artifact
ondo3CLAIM_CAPTURED 1 · CLAIM_MEASURED 1 · UNMEASURED 12026-09-222026-09-22UNSCHEDULEDartifact

Disclosed about our own records

  • claimreg-ondo-chainlink-2026-09-22 — quoted: “UNSIGNED — signer ceremony pending; OTS-anchored instead (no key required)”. That sentence is the source registry's own wording, quoted. This register does not confirm anchoring for it: a receipt file is SUBMITTED until its upgrade has been run and its attestation path verified. The receipt's state here is RECEIPT_PRESENT_UPGRADE_UNVERIFIED (spec 8.1-8.3).

Machine-readable: /api/claims/register · static bytes. Filter with ?subject= or ?state=.

Run it yourself — including against us

A specification with no runnable implementation is a manifesto. This one captures a claim from any public page, extracts the visible text by the rules in §6.3, computes the digests, writes a conforming artifact, and verifies one. It has no dependencies. The verifier rejects an artifact whose bytes have been altered after the digest was taken — including a rewritten evidence URL, which is the field that has historically sat outside the hash.

curl -fL https://councilof.ai/spec/claim-maintenance/v0.2/reference/claim-capture.mjs -o claim-capture.mjs

node claim-capture.mjs --url https://example.com/ \
  --subject "Example Corp" --identifier example.com --identifier-kind domain \
  --claim "market leader powering the majority of the sector" \
  --plan "majority = >50%; capture weekly against the public breakdown" \
  --out artifact.json

node claim-capture.mjs --verify artifact.json

Verification answers whether these bytes are internally consistent and conform to the state machine. It never answers whether a claim is true, and a sound signature over a measurement performed wrongly is a sound signature over a wrong number.

Artifact schema: claim-artifact-v0.1.schema.json · Specification as Markdown: claim-maintenance-v0.1.md · All versions: /spec/claim-maintenance/

Where the rest of the record lives

  • The corrections ledger — what we got wrong, how it was caught, and the fix, dated. A body that observes other people's claims and does not publish its own errors is asking to be trusted rather than checked.
  • Our own claims register — the claims we make about our own capabilities, with the status each has earned. A separate document from the register above, and deliberately so: one is what we say about ourselves, the other is what we observe about others.
  • How we measure — the measurement discipline this specification borrows its vocabulary from. We measure; we do not certify.

If a record here does not match what you can check, that is a defect and we want it. Write to nicholas@csoai.org; corrections are published at /api/corrections, where anyone can read them.