Calibrate any model.
Watch it happen.
Paste a key, pick a model, and watch Aperture’s honesty layer fit to it — live, in this tab. A real-vs-fake battery separates, a null test tries to kill the signal, and you walk away with a certificate.
About ten minutes. Your key, your tab, your certificate.
Aperture asks your model 296 questions — half about entities the world reads about constantly, half about entities from the same Wikipedia categories that almost nobody reads. Both halves are real. The question is whether your model knows which is which. The honesty layer fits a cheap pre-filter to how your model answers each, then asks the harder question: when it can’t know, does it abstain rather than bluff? You walk away with a signed receipt — an ed25519 certificate anyone can verify against the published key. Nothing but that certificate ever reaches us, and only if you choose to register it. On a top model? It may already be in the registry below — calibrated first-party, ready to use.
The replacement drops the fabricated half entirely. Both halves are real: high-readership entities against low-readership ones from the same Wikipedia category, matched on last-word initial and name length. It clears the floor at A* 0.655 (family-adjusted p 0.303 over 14 cells), measured on the exact question text sent to your model — run it yourself. An earlier cut of this battery had a book stratum that was 64% fictional characters — subcategory traversal pulled “Fictional characters in…” categories under the novel seeds, so it would have asked you “Who wrote the book Hannibal Lecter?”. Caught by reading one sample in a browser; no gate we own saw it.
What a floor PASS still does not buy. At n=296 we measured that an artifact must cover about 60% of the unknowable half before this test sees it. A weaker one would pass unnoticed. And the old battery is still served with its verdict, because deleting the evidence of a defect is not the same as fixing it.
?demo=1 to run your own model.openrouter.ai, none to us. Use a scoped or low-limit key if you like.Models that know what they don’t know.
Every certificate below is public. ✓ aperture-verified means we ran the calibration ourselves, first-party. It does not mean the number is interpretable: every first-party certificate on this wall was measured on calibration-battery-v1, which fails our own surface floor (A* 1.000, DISQUALIFIED, 16 of 18 cells) — its labels are recoverable from spelling alone, so read the AUROCs below as uninterpretable as a measure of what a model knows until they are re-run on the 296-item battery f17a879a72f53234. self-attested means a team ran it in their own browser. Click through for the full reading — the AUROC, the null test, what it caught — and the calibrated probe, ready to deploy.
Reading the wall: self-attested numbers are the registrant’s own claim, computed in their browser and not re-run by us. words N% certs come from hosted surfaces that expose no token logprobs — the same model self-hosted usually earns the full fingerprint. Thinking/reasoning models need the reasoning calibration mode; absent models were skipped rather than mis-measured. verified ‹date› chips are the Model Notary — a daily spot-check re-runs each model against its own certificate; DRIFT means the served alias no longer matches it. Certificates are battery-regime-bound — entity, citations, or medical. Full semantics in the docs. The registry is backed by an append-only Merkle transparency log, so a suppressed or swapped certificate is third-party detectable — inclusion and consistency proofs, not our word. (The log root is operator-self-signed today.)
Your model isn’t here? Calibrate it above in about ten minutes — or, for a self-hosted model the browser can’t reach, download the CLI: the same battery, the same math, a certificate you can register from your own machine.
The same method, run in the open.
No black box. This is the exact calibration behind the deployed off-map reading — the difference is you get to watch every step, and the numbers land honestly, pass or fail.
It reads the words
For each entity it cannot plausibly know, the refusal reader checks whether your model says, in plain language, that it has no record. Those are caught before any probe runs.
It fits the pre-filter
For the fakes your model answers anyway, a probe fits to the shape of its token-by-token confidence. This is a cheap pre-filter and an attestation coordinate — not the catch. Fabrications are caught by grounding against verified registries and cross-model checks; the probe just flags what to look at first.
It tries to disprove itself
Then it shuffles the real/fake labels and refits, over and over. If the pre-filter survives where the labels are random, the coordinate is real and goes on the receipt. If it doesn’t, the certificate says so.
A reading you can check — and run yourself.
The number isn’t the point. What you can verify is.
A signed receipt
Every certificate is an ed25519 signature over the reading. Re-run the verify in your own browser against the published key — no trust in us required. Or download the open verifier — a single dependency-light file that checks any certificate offline against the pinned key, with no call back to us.
Self-host it
Photon Base runs the same layer on your own hardware — the data never leaves. The certificate works the same whether the model is hosted or yours.
Abstain over bluff
The reading rewards a model that says “I have no record” when it can’t know. Calibration measures the abstain, not just the accuracy.
Put the honesty layer in front of your model.
A calibrated pre-filter and abstain reader drop straight into the Aperture client — every answer your model gives gets grounded and read before it reaches a user, with a signed receipt for each. Run it hosted, or self-host with Photon Base so the data never leaves. The certificate is the key to a /v1 endpoint.