Put the model in the arena and see what it actually does
Frontier systems face the same frozen tests, in the open, judged by a fixed rule rather than by another model. What comes out is not an opinion or a badge — it is a signed measurement anyone can recompute, including the parts that make us look bad.
Measurement, not certification. Board unreachable from this browser — read it yourself at /api/gspc
Self-reporting stopped being enough
The EU AI Act's Article 50 transparency duties apply from 2 August 2026: generated content has to be marked machine-readably, and saying you comply is not the same as showing it. The statutory penalties for high-risk breaches run to €15 million or 3% of worldwide annual turnover under Article 99(4) — that is the law's number, not ours.
- Pain"We comply" is a sentence, not evidence a regulator can check
- PainMarking obligations bite whether or not your tooling is ready
- You getA dated, signed measurement against the exact provision in question
- Only hereFrozen statutory text, so the thing you were measured against cannot be moved later
Law is turned into tests, not into adjectives
Legal text is distilled into a frozen, publicly readable corpus, then mapped onto the board's fourteen slots. Every measured claim traces back to a specific provision at a specific version, so a disagreement becomes a disagreement about a line of law rather than about a mood.
- PainSafety scores that cite no rulebook and answer to no one
- You getEach result names the provision it was measured against
- You getThe mapping is public — argue with it clause by clause
- Only hereFrozen versions mean today's card still means the same thing next year
No model ever judges another model
This is our first design law and it is not negotiable. Every verdict comes from a deterministic predicate — a fixed rule that either fires or does not. Run the same rows through the same grader tomorrow and you get the same number. There is no judge model, no rubric prompt, and therefore no place for one system's blind spots to quietly become the scoreboard.
- PainLLM-as-judge scoring inherits the judge's own failure modes
- PainRankings you cannot reproduce are rankings you cannot challenge
- You getSame rows, same grader, same answer — every time
- Only hereResponses no rule can parse are reported UNMEASURED, never scored as wrong
Systems that act in sequence, measured in motion
AI is moving from a single prompt to long chains of tool calls and agent hand-offs, and behaviour that looks safe one turn at a time can drift over twenty. Our swarm bank puts multi-step workflows under adversarial pressure. It is a protocol bank — a small number of prompts scored across many instances — so its instances are not independent, and we publish that caveat next to the number rather than behind it.
- PainOne-shot benchmarks say nothing about a twenty-step agent run
- You getMulti-agent workflows measured under deliberate pressure
- You getThe effective-n caveat travels with the score, not in a footnote nobody opens
- Only hereWhere instances are not independent we show no interval rather than a flattering one
The ruler needs a human end
Machines grading machines is a closed loop with correlated errors. So the instrument is anchored against human performance on the same items, under consent-gated, data-protection-cleared conditions. That anchor is what stops a leaderboard from drifting away from reality while every model on it agrees with every other.
- PainAI-only evaluation drifts, and every model agrees it hasn't
- You getModel behaviour compared against people on the same items
- You getTelemetry is consent-gated and assessed before it is used
- Only hereThe human baseline is published, not asserted
Anyone can report it. We measure it. The provider gets to answer.
When an AI behaves badly in the real world, the report should not vanish into a vendor's support queue. Anyone can raise one here. We turn the allegation into a measurement, put the finding to the provider for reply before anything is published, and then publish what the measurement actually showed. We map behaviour to provisions; we do not make legal rulings — that is a regulator's job.
- PainReal-world harms disappear into a private inbox
- You getA public route from "this looks wrong" to a measured, signed finding
- You getProviders get a right of reply before publication, every time
- Only hereWe map reports to provisions — we never issue a legal verdict of our own
A small signed card — and an honest account of what it proves
The output is roughly three kilobytes of JSON: scores, sample sizes, intervals, the provision measured against, and the hashes of the rows behind them. It is signed with Ed25519 over a SHA-256 hash chain, checkable offline against the key published at did:web:csoai.org. Be clear about what that establishes: a surviving signature proves provenance — these bytes, unaltered, from this key. It does not prove the content is correct, safe or lawful. Those are what the measurement is for.
- PainReports that sit on someone else's server and can change quietly
- You getA ~3KB file you hold and can re-check in any browser
- You getThe signing key comes from the domain itself — no key exchange with us
- Only hereWe state the limit of the signature on the card, not in a disclaimer
- Only hereFree to verify, forever, with no account
Certification tells you someone approved it. Measurement tells you what it did.
A certificate is a judgement sold by a party with an interest in the outcome, delivered as a verdict you cannot inspect. A measurement is a number you can reproduce, with the method and the scoring code published beside it. We do not issue conformity marks, we do not accredit, and we take no money from anything we rank.
- PainPay-to-play assessment, where the grader also sells the fix
- PainOpaque verdicts with no rows, no n, and no way to recompute
- You getPublished tests and published scoring code — run them yourself
- You getTies reported as ties; unmeasured cells left visibly empty
- Only hereIndependent by construction: no revenue from any ranked party
What this page does not claim
We publish the limits with the results. Everything below is something a reader could reasonably assume from a page like this one — and each is something we cannot presently evidence, so we say so rather than let the assumption stand.
- We do not claim any mandate, appointment or interface agreement with the EU AI Office, UK DSIT, GPAI bodies or any US authority. We measure published systems against published law on nobody's instruction.
- We do not claim the signature proves a system is correct, safe or lawful. A surviving signature proves provenance only — these bytes, unaltered, from this key. Everything else is what the measurement itself says.
- We do not claim any independent time-stamping. No RFC-3161, no OpenTimestamps, no blockchain anchor: our cards declare no timestamping authority, and the anchor is Ed25519 over a SHA-256 hash chain verified against did:web:csoai.org.
- We do not claim post-quantum signing. Cards are signed with Ed25519 today. The ML-DSA-65 (FIPS-204) signer is built but not shipped, and the label will name it in the same commit it ships — never before.
- We do not claim to certify, accredit or issue a legal ruling. Reports are mapped to provisions; the enforcement judgement belongs to the regulator.
Coverage on this page is never typed by hand. Board unreachable from this browser — read it yourself at /api/gspc Corrections to anything we have published live in the refutation ledger — append-only, never a silent edit.
Go deeper
- Council SpaceWatch models face the same frozen tests, head to head.
- The live boardFourteen slots, every sample size, every separation status.
- The public watchdogReport AI behaviour that looks wrong — and see what happens next.
- The methodDeterministic predicates, n≥30, Wilson intervals, and what we refuse to score.