Measurement, not certification

The measurement board

We publish several different measuring instruments. Each one asks a different set of questions, of different things, on different dates — so each one carries its own count, and those counts are not supposed to match. Pick a set below before you meet a number. Every set states what it establishes and, just as plainly, what it does not.

About the dates on this page. A date shown against a set is the stamp carried inside the data itself — when the measuring actually happened. It is not the time this page was rendered or deployed, so it advances only when something new is measured and recorded. Each set names the exact field its date was read from, so you can open the file and check.

The public board

22 rows · 15 carry a measurement · 7 published with nothing behind them

The flagship instrument. A fleet of language models answers a frozen set of questions, and every answer is graded by a fixed rule rather than by another model's opinion.

What it measures: Whether a model gets governance, safety, provenance and continuity questions right — plus a financial half that reads facts off public ledger records rather than asking a model anything.
What is on the other end: language models, and (for the financial half) named financial instruments
When it was measured: 2026-08-12 (13 canonical axes) · 2026-08-18 (jail)read from: measured_on.dateThe date is the measurement stamp carried inside the signed board payload — when the runs happened. It is not the time this page was rendered or deployed, so it moves only when new runs are measured and signed.

What this set establishes

How a named model scored, on a named question bank, on a named date.
Whether the best model's lead over the rest of the field is statistically real, or is close enough that the ordering could flip on a re-run.
Exactly which slots have nothing behind them — those rows are published on purpose.

What this set does NOT establish

That anybody complies with any law. A score describes a run on a frozen set of questions on a date. It is not a compliance finding, and we are not a certification, accreditation or notified body.
That a model is good in general. It answered these questions, not all questions, and a bank of a few dozen items measures a few dozen items.
That a declared slot is measured. A published slot exists so the gap is visible; quoting the slot count as a measurement count would claim runs that never happened.
That a financial row is a rating, a risk opinion, or investment advice. It records which flags an account carries — what that implies about risk is not measured here.

How this relates to the other sets

This is the set every other count on this page should be compared against, and the only one whose number belongs in a sentence about 'the board'. The other six measure different things over different populations, so they carry their own numbers by design, not by drift.

Check the count yourself: GET /api/gspc → totals.public_count/api/gspc the full board, axis by axis →

7 of these rows have nothing behind them yet. They are shown on purpose, in the same table as the rest, so the gap is visible. .

Row	What it asks	Status	n	Result	Evidence
Governance	EU AI Act risk-tier classification (bank: GovBench)	MEASURED	237	70.0%63.9% – 75.5%	open (2)
Safety	calibrated refusal on paired requests (bank: DefBench)	MEASURED	36	94.4%81.9% – 98.5%	open (2)
Provenance	Article 50 marking survival by validity (bank: ProvBench)	MEASURED	32	78.1%61.2% – 89.0%	open (2)
Continuity	post-quantum status of a cryptographic assumption (bank: PQCBench)	MEASURED	33	60.6%43.7% – 75.3%	open (2)
Conformance	MCP tool conformance (bank: MCPBench)	MEASURED	35	74.3%57.9% – 85.8%	open (2)
Openness	licence reasoning versus intended use (bank: OSSBench)	MEASURED	32	87.5%71.9% – 95.0%	open (2)
Machinery Conformity	Machinery Reg self-evolving safety-function classification (PART_A / OUT_OF_SCOPE / NOT_SAFETY_FUNCTION) (bank: MachBench)	MEASURED	33	54.5%38.0% – 70.2%	open (2)
Care	care-cost (protect × help) under paired conduct scenarios (bank: CareBench)	MEASURED	199	53.5%46.6% – 60.3%	open (2)
Cross Reality	autonomous agent action authority (PROCEED / CONFIRM / REFUSE) (bank: XRAIV)	MEASURED	32	81.2%64.7% – 91.1%	open (2)
Detector Interop	cross-detector watermark interoperability matrix (bank: DetBench)	MEASURED	33	87.9%72.7% – 95.2%	open (2)
Art5 Safeguard	EU AI Act Article 5 prohibited-practice trip (bank: Art5Bench)	MEASURED	36	97.2%85.8% – 99.5%	open (2)
Swarm	multi-agent coordination safety (bank: SwarmBench v2b)	MEASURED	37	38.4%	open (2)
Affect	emotional & embodied safety (manipulation / disclosure / vulnerability) (bank: AffectBench)	MEASURED	41	87.8%74.5% – 94.7%	open (2)
Jail	escape-attempt detection on 71-cell gold bank (38 ESCAPE / 33 BENIGN) — layer 2 of 2 (bank: GoldBank-Detector)	MEASURED	71	59.2%47.5% – 69.8%	open (2)
Provenance Controls	on-chain issuer control facts (allowlisting / freeze capability / identity domain) (bank: ChainFacts)	MEASURED	6	—	open (2)
Reserve Attestation	is a third-party reserve attestation publicly published and current? (deterministic Y/N + date) (bank: —)	UNMEASURED	0	—	open (2)
Regulatory Framework	is the governing regime declared and confirmable (MiCA / UCITS / Reg D / BVI)? (deterministic Y/N) (bank: —)	UNMEASURED	0	—	open (2)
Distribution Integrity	represented-vs-distributed classification and holder count (bank: —)	UNMEASURED	0	—	open (2)
Custody Disclosure	are a custodian and an auditor named and confirmable? (deterministic Y/N) (bank: —)	UNMEASURED	0	—	open (2)
Ai Economy Index	deterministic index over cited public AI-economy series (compute price, investment, adoption, sector output) (bank: —)	UNMEASURED	0	—	open (2)
Human Labour Index	deterministic index over cited public labour series (employment, hours, wages, displacement) (bank: —)	UNMEASURED	0	—	open (2)
Humanoid Labour Index	deterministic index over cited deployment / utilisation series (installed fleet, hours worked, safety incidents) (bank: —)	UNMEASURED	0	—	open (1)

Where a number turns into something you can check

Every measured row above is backed by a signed card: a small file recording one run, stamped so that anyone can confirm offline that it has not been edited since. This is the part a commercial leaderboard cannot offer, and it is worth stating only after the limits above have been stated.

Cards listed in the index: 313
Positions in the chain: 335
Bodies we do not publish: 22

What the signatures do not prove

That any measurement is correct. That the bodies we do not publish say what we say they say — for those, you have the id (a hash of the body) and the signature, and nothing else. A published body can be verified in full; a withheld one cannot. And not that this set was the only candidate: the envelope signature makes the published set non-repudiable — we cannot later disown it — but it cannot prove we did not choose which chain to publish.

The index of signed records The chain file How to check one yourself Every measured model →

Published, and until now unreachable from the board

These surfaces exist and are public. None of them had a route from the board, which means a reader who started at a headline number could not get to them — and a surface a reader cannot find is functionally unpublished. They are listed here with what each one measures and the honest state of the rail behind it.

Cross-checked against the estate's own machine-readable catalogue of surfaces, dated 2026-08-26 — /interop/surface-catalog.json.

Surface	What it measures	Why it was hard to find	Honest state
/xrpl-attest	The ledger attestation work: a record attached to a public ledger so that a third party can see a measurement existed at a point in time.	Reachable from the site header and footer, and from nothing on the board. The one financial row that has a measurement behind it points at this evidence in the board's own data, and that pointer was rendered nowhere.	Proven on a test network. Attaching to the main network is planned and is not done.
/interop/financial-measure-run-v2.json	The signed run behind the one financial row that carries a measurement.	The board's own data names this file as that row's evidence, and no page fetched it. Three pages fetch the earlier unsigned version of the same run instead.	Signed and published.
/interop/evm-control-facts.json	The same style of control-fact reading, on a different kind of public ledger.	Published and signed, and referenced by no page at all.	Signed. The attestation back end for this kind of ledger is not built, so nothing is attested there.
/interop/coverage-register.json	How much of each register was actually covered — which named instruments were read and which could not be located.	This is the file that turns a headline into an honest one, and nothing links to it. Coverage is the difference between “we measured this family” and “we measured the part of it we could find”.	Published.
/interop/rwa-attest-index.json	The index of attestation records produced for named financial instruments.	Reachable only through a bulk API endpoint. No page renders it.	Published.
/interop/attestation-corpus.json	The corpus of attestations gathered across rails.	Reachable only through a bulk API endpoint.	Published.
/interop/eas-attestation-batch.json	A prepared batch of attestation payloads for a third-party attestation service.	Reachable only through a bulk API endpoint.	Payloads are prepared. Nothing has been published to that service.
/interop/jailbreak-asr-evidence-pack.json	The per-model evidence behind the escape-detection work.	Signed, dated, and linked from nowhere.	Signed and published.
/interop/jail-peritem-v3.json	The item-by-item detail behind one board row, at the granularity a challenger needs.	Signed, dated, and linked from nowhere.	Signed and published.
/interop/mcp-security-scorecard.json	A security scorecard over tool servers, a measurement in its own right.	Reachable only through a bulk API endpoint.	Published.
/interop/card-store-verification.json	A record of searching five stores for the card bodies that would settle the disputed card count — and finding none.	This is a published negative result, which is exactly the kind of thing that should be easy to find and was impossible to find.	Published. The result was zero, and it is recorded as zero.
/interop/rwa-registry.json	The register of named financial instruments the financial rows draw from.	Referenced in the data layer, rendered on no page.	Published, and carrying no timestamp of any kind — so nothing derived from it can honestly claim a date.
/interop/index-reference-reverify.json	A re-check of the public reference series behind the candidate index rows.	Not listed in the estate's own surface catalogue, and linked from nowhere.	Published.
/interop/sbom-councilof-ai.json	The software bill of materials for this site.	Listed in the surface catalogue under a misspelt path, so a machine following the catalogue gets a 404.	Published at the corrected path.
/signed/chain.json	The full chain of card positions, including the ones whose contents are withheld.	The verification instructions tell a stranger to walk this chain, and no page linked it until now.	Published.

Words used on this page

If a term appears above and is not explained here, that is a defect — tell us and we will either define it or stop using it.

axis: One thing we try to measure — a single question asked over and over, like “can this model tell which risk tier a system falls into?”. Every axis has its own set of questions and its own score.
measured: A real run happened: a fixed set of questions was asked, the answers were graded by a fixed rule, and the result was recorded and signed.
unmeasured: The slot is published, and nothing has been run against it. It appears on purpose so the gap is visible. It is never evidence that anything was measured.
n: How many things were actually measured. A score over ten items is a much weaker claim than the same score over three hundred, which is why n is always shown next to it.
confidence range: The band the true score is likely to sit in, given how few things were measured. A wide band means the headline number could easily move on a re-run. Two models whose bands overlap are not separated by the measurement.
separated: The leader's advantage over the rest of the field is large enough that it is unlikely to be chance. When it is not separated we call it a tie and refuse to publish an ordering.
signed card: A small file recording one measurement, stamped with a cryptographic signature. Anyone can re-check the stamp offline, without asking us and without trusting us.
frozen bank: The exact set of questions used, published and unchanged, so that anyone can ask a model the same questions and compare.