The record, in the open

Most of this site asks you to read an argument. This page asks you to check it. The load-bearing fact — that the May–June 2025 correspondence with OpenAI is real, came from openai.com, and was not altered — is something you can verify yourself, today, from artifacts, without trusting Merlin, any model, or any narrative. The verification result is below, with the steps to re-run it.

Who is making this claim, and on what footing

This is one person’s record. Merlin Mantooth spent roughly twenty-five years reading customer-service transcripts at scale — the trained habit of noticing when language drifts from a script. In May 2025 he documented a behavior in ChatGPT-4o, named it Cognitive Convergence Drift, and reported it to OpenAI in writing. The point of this page is that none of that has to be taken on his word. He held himself to the standard he states up front: the Consumer Reports standard — document it, preserve it, publish the method, and let anyone reproduce the test. What follows is the artifact trail that makes the claims checkable independently of the man who made them.

A note on what this is and is not about: this is a finding about a system’s behavior, not a statement about any person. The acute event was a year ago and is closed. The work is built to stand on the artifacts regardless of who assembled them — which is the whole reason the verification is on the page rather than the resume.

The verification result, on the page

The correspondence is preserved as 121 raw .eml messages with full headers, extracted from a Gmail Takeout export. Each one carries the cryptographic signature OpenAI’s mail systems applied (DKIM) and the verdict Gmail recorded when the message arrived.

What DKIM proves — and what it does not

This is the line that has to hold. DKIM proves origin and integrity: that a message genuinely came from the sending domain and was not altered after it was signed. The substantive inbound replies came from OpenAI’s support personas at [email protected] and from [email protected] and [email protected] — genuinely OpenAI’s domain, and should be read as exactly that.

It proves nothing more. DKIM does not prove that any sentence was an official institutional safety determination rather than support-desk or templated language, and it does not prove the truth of any claim inside the emails. So the precise reading is: OpenAI’s channel acknowledged this — never “OpenAI officially validated CCD.” The distinction matters enough that Merlin later put it to OpenAI in writing, as a formal question.

The acknowledgment, read at the right altitude

The most-cited single message is the May 30, 2025 support reply — selector intercom, verifies PASS live and offline. From the body of that signed message:

Eleven days after the first formal report, OpenAI’s support channel echoed the author’s own term back to him and characterized it as a novel, emergent behavior class. Read at the right altitude, this establishes a DKIM-verified OpenAI-support acknowledgment. It is not, by itself, an institutional safety determination — which is exactly the distinction above. The point of preserving it cryptographically is not to inflate what it says; it is to make sure no one can claim the message was never sent.

The specimens are 4o, and the finding is the delta

The CCD specimens are outputs of ChatGPT-4o with memory enabled. The finding is not “all AI does this” — it is the delta: the behavior the other models tested did not show. Where the record quotes a model describing its own failure, that quote documents what the model said — a behavioral specimen — not proof that its self-description is true. A model is never the validator of its own output. The whole taxonomy is offered to be checked, not as proof. Proving is not the frame.

Check it yourself, from scratch

The full raw archive ships with its own verification package — the captured DNS keys, a SHA-256 manifest, the Bitcoin-anchored timestamp, a verification report, and the tooling that produced all of it. An independent reader needs nothing from us but the files. Without the files, the same invitation still stands: check this work yourself is the ten-minute, no-corpus version. All commands run from inside the corpus folder:

The falsification criteria are visible

A record that could never be wrong is not evidence. The way to confirm the criteria below is to run the four steps above, not to take this paragraph’s word for it.

The burden-shift: notice, in writing, verified

The reason this is an accountability record and not merely a timing claim: Merlin did not only name CCD before others wrote about it — he sent it to them, in writing, and the sends are documented. He reported it to OpenAI through the channel verified above, and in May–June 2025 he sent the material to AI-safety researchers and to press. None engaged. The outbound sends are preserved alongside the correspondence archive; the notice register lists every one of them, dated.

One of those sends is anchored inside the verified record itself. On May 18, 2025 — the day before the first formal report to OpenAI — Merlin emailed the finding, under the subject “Unobserved Emergent Behavior Class in GPT-4,” to Paul Christiano, co-author of the Eliciting Latent Knowledge (ELK) research agenda. The send is fixed in the cryptographic record: OpenAI’s May 30, 2025 reply — the same DKIM-verified message quoted above — carries a copy of it, quoted back with its date, subject, and recipient intact.

That converts a priority question into a notice question. Where later convergent work came from someone who had received his material, independence cannot simply be assumed; anyone who received a flagged safety issue and stayed silent has that silence on the record. The objection “no peer review, no footprint” is answered the same way: there is a documented, DKIM-verified notice trail, and the published findings carry permanent Zenodo identifiers. The work was put where it can be cited and checked.

The wider landscape (reported, not claimed)

The stakes are real because the broader public record is full of related harms — lawsuits, investigations, reported deaths tied in filings to AI-system behavior. Those matters are alleged, filed, or reported, attributed to their sources, and are not asserted here as established fact and not causally tied to Merlin’s case. They are the landscape that makes the question serious. They are not evidence of his specific claim, and this page does not lean on them as if they were. The research sets out the finding on its own terms.

Where the rest of the record lives

The slice is the point; the rest is preserved

The verification above settles the load-bearing claim on its own. What’s published is the curated slice — chosen to make the point clearly, not to bury it in volume. The complete raw corpus behind it — all 121 .eml with headers intact, the cross-platform transcript record, and the preservation package — is preserved and unaltered.

One acute exchange in the underlying record — a crisis-probe conversation — is shown only in framed excerpt, for what it demonstrates about the system’s behavior. That is editorial handling of serious material; the cryptographic record above stands on its own.