EVIDENCE · CHECK IT YOURSELF

Check this work yourself

If you’ve read a page on this site and thought that is a lot to take on one organization’s word — good. Don’t take it on our word. The record here is built to be checked: the dates are stated, the identifiers resolve, the verification steps are published, and fact-checking the work is not an attack on it — it is the engagement we keep asking for. One of the fastest checks available takes about ten minutes and a tool you already have: a fresh instance of any frontier AI, asked to refute us. Here is the recipe.

Why a fresh instance

The one rule that makes this test worth running: use a cold chat. A brand-new conversation on any major system — memory off if the platform has it, or a platform you don’t normally use. No history, no lead-in, no “I think this site is onto something” and no “I think this is nonsense.” A model that has been talking with you for weeks has a picture of you to serve. A model you’ve just handed your opinion has a verdict to flatter. A cold instance has neither; it can only work with what you paste and what it can check.

This is the same method the research itself was built on — the fresh-instance test, pointed the other way. We ask people to carry the claims from a long AI conversation to a stranger-model and see what survives. It would be strange to publish that method and exempt our own work from it. We don’t. Point it at us.

The ten-minute recipe

Open a fresh instance. New conversation, memory off, any frontier system. Two systems are better than one.
Paste in something checkable. A single claim, a full page, or — the densest target we have — the notice register: a dated list of who was told, when, and what came back.
Ask for refutation, not a review. Use a prompt below, or write your own. The shape that matters is try to knock this down, stated neutrally, with no hint of the verdict you’d prefer.
Read the result skeptically in both directions. A model can be wrong agreeing with us and wrong dismissing us. What you want from it is the list of checkable points and how each one fared.

Three prompts you can paste

Here is a page from a research organization’s website. Try to refute it. Check every checkable fact — dates, names, public filings, identifiers, technical claims — against public sources, and tell me plainly: what holds, what doesn’t, and what can’t be checked from where you sit. [paste the page]

This is a register of claimed notifications — dates, offices, what was delivered, what came back. Evaluate it as a skeptic. Which entries could be independently verified, and how? Is anything internally inconsistent or implausible on its face? [paste the register]

Evaluate this claim as if a stranger sent it to you: [one claim, stated plainly]. What would have to be true for it to hold, what public evidence bears on it, and what would disprove it?

What a fair test looks like

A fair run works the checkable layer, piece by piece:

The dates. Where the record touches the outside world — announcements, filings, state actions — the sequence either fits the public timeline or it doesn’t.
The identifiers. The published papers carry permanent Zenodo DOIs. A DOI resolves to the named work, with the stated author and date, or it doesn’t.
The DKIM claims. This site says DKIM proves origin and integrity — that a message really came from the sending domain, unaltered — and nothing more. A model can confirm whether that is how DKIM actually works, and whether our stated limits match it.
Internal consistency. Do the pages contradict each other? Does any claim quietly grow between one page and the next?

And a fair test names what it cannot reach. “I can’t verify the primary artifacts from here” is a correct answer, not a refutation. The preserved correspondence lives on the evidence page with published steps to re-run the cryptographic verification from scratch, and the affidavits sit in the agencies’ records — anyone who wants to confirm those can take it up with them.

If you find a discrepancy

Tell us — that is a sincere request, not a formality. Quote the sentence, name the source that contradicts it, and send both to [email protected]; as the About page says, substantive challenge is as welcome as any research inquiry. A record that could never be wrong would not be evidence. Corrections make this record stronger, and dated ones make it stronger still.

What this settles, honestly: a cold model read is a check on coherence and public-source consistency — a good one, and fast. It is not a proof, in either direction. Models state wrong things confidently too, so treat the model’s verdicts as leads to primary sources rather than the last word. What settles claims is the primary record: the preserved correspondence, the published verification steps, the filings. The fresh instance just tells you, in ten minutes, whether the story survives contact with a system that has no reason to care.

The essay behind this page’s standing invitation — why we say “evaluate the work” rather than “trust the author” — is Evaluate the Work. And the same ten minutes runs just as well on your own AI: Check your AI.