On Thursday, July 10, OpenAI dropped a PDF on its content delivery network. The document, attributed entirely to GPT-5.6 Sol Ultra, purports to prove the Cycle Double Cover Conjecture, a graph-theory problem that has resisted solution since the 1970s. Within an hour, the claim was on the front page of Hacker News. Within two, someone had edited the conjecture’s Wikipedia entry to note that “OpenAI company claimed the problem was solved.”

The model, we are told, used 64 subagents working in parallel and produced the proof in under an hour. Ethan Knight, an OpenAI executive, posted the prompt and the PDF on X, calling it something the company was “excited to see what you all do with.” The framing was unmistakable: this was a product launch, not a research submission.

And that is the real story. Not that a large language model may have cracked a hard math problem—though that would be remarkable if it holds up. The story is that OpenAI chose to announce it not through a journal, not through a preprint server with an accompanying human paper, not even through a press release that named the researchers who designed the system. It announced it by posting a PDF on a CDN and letting social media do the rest.

The Proof Is in the Posting

Mathematics has a well-worn path for claims of this magnitude. You write the proof. You circulate it among colleagues. You submit it to a journal. Reviewers—actual humans who have spent careers in the field—interrogate it for months. Sometimes a flaw is found. Sometimes the proof is correct but the attribution is messy; credit gets negotiated. The process is slow, adversarial, and imperfect. It is also the only mechanism the discipline has for separating signal from noise.

OpenAI skipped every step. The PDF did not appear on arXiv. No mathematician was listed as a co-author or a verifier. The company’s own blog post, if one exists, was secondary to the artifact itself: a raw claim, served from a corporate CDN, with the model named as the sole author. This is not how you contribute to human knowledge. This is how you assert authority over it.

A postdoc in a math department Slack channel, after the PDF began circulating Thursday evening, put it plainly: “We’re not even sure if it compiles in Lean, and they’re already editing Wikipedia.”

Who Verifies the Verifier?

None of this is to say the proof is wrong. It may be elegant. It may be a genuine breakthrough. The conjecture—that every bridgeless graph admits a family of cycles covering each edge exactly twice—has attracted some of the best minds in combinatorics. If GPT-5.6 Sol Ultra really did produce a valid proof, it would be a staggering demonstration of what these systems can do when pointed at a well-defined formal problem.

But we do not know that yet. What we know is that OpenAI, a company that sells access to models, has claimed a result that, if accepted, would redound entirely to the model’s reputation. There is no independent verification. There is no named human accountable for the claim. There is only the PDF, the CDN, and the viral thread.

This is a pattern that should worry anyone who cares about the integrity of expert knowledge. When a pharmaceutical company announces a drug trial result, we expect the data to be published and scrutinized—not posted on a corporate server with a “see for yourself” shrug. When a bridge engineer certifies a design, we expect a stamp from a licensed professional, not a simulation output attributed to the software. The norms exist because the stakes are high, and because institutions, for all their flaws, provide a check on the impulse to announce first and verify later.

The Real Conjecture Is About Authority

The Cycle Double Cover Conjecture is, in one sense, a perfect test case for this kind of maneuver. It is famous within a subfield but obscure to the general public. Few reporters can assess the proof. Fewer still will try. The result is a vacuum that OpenAI can fill with its own framing: “AI solves 50-year-old math problem.” The headline writes itself, and the question of whether the proof is actually correct becomes a secondary concern—something for the mathematicians to sort out later, after the market has already priced in the achievement.

This is not a column about whether AI can do mathematics. It can, increasingly, and that is a genuine intellectual achievement. This is a column about who gets to declare that a problem has been solved, and on what authority. For half a century, that authority rested with a distributed community of researchers, journals, and conferences—a community that moved slowly, argued fiercely, and demanded rigor. OpenAI is betting that a PDF on a CDN, amplified by enough retweets, can substitute for all of that.

Maybe the proof checks out. If it does, the mathematicians who verify it deserve the credit for the verification, not the model for the output. If it doesn’t, the episode will be remembered as a stunt. Either way, the method of the announcement reveals something about the moment we are in: the institutions that once mediated between a claim and its acceptance are being treated as optional, and the companies building these systems are happy to fill the gap with their own distribution channels.

A proof is not a proof until it has been read, understood, and certified by people who can be wrong. That is the point of the process. Skipping it does not make you faster. It makes you a press release.

Sources