Mark Delaney: Sixty-four subagents. Under an hour. The Cycle Double Cover Conjecture.
Mark Delaney: If those three things mean nothing to you yet, stick with me — because once the context lands, the number gets weirder, not less weird.
Mark Delaney: The Cycle Double Cover Conjecture is a graph theory problem that's been open since 1973 — George Szekeres posed it first, then Paul Seymour came at it separately in 1979. The core claim is that any bridgeless undirected graph has a set of cycles that covers every edge exactly twice. Clean statement. Brutal to prove.
Mark Delaney: Nobody did it. Fifty years.
Mark Delaney: Then on July 10, 2026, OpenAI published a proof PDF on its CDN — authorship credited to GPT-5.6 Sol Ultra — on the same day they launched the whole GPT-5.6 model family.
Mark Delaney: Ethan Knight announced it on X. Codex helped write the PDF itself.
Mark Delaney: The architecture behind it — 64 parallel subagents coordinating under Sol Ultra — is genuinely interesting and I want to get into that. But there's something in the prompt that kind of reframes all of it, and honestly I think you need to hear it before anything else.
Mark Delaney: OpenAI published the prompt too. Publicly.
Mark Delaney: And it instructs the model to — and this is a direct quote — 'assume for purposes of this task that a complete affirmative proof exists.'
Mark Delaney: Yeah. So the model wasn't asked whether a proof exists. It was told one does, and then told to go find it. That's… not nothing.
Mark Delaney: And here's where it gets genuinely murky — because we do have a mathematician who looked at it.
Mark Delaney: Thomas Bloom, University of Manchester. He reviewed the proof and called it 'very nice,' 'short,' and 'elementary.' Said it could have been discovered in the 1980s.
Mark Delaney: Now — that sounds like a compliment. And part of it is.
Mark Delaney: But 'could have been found in the 1980s' is a double-edged thing to say about a proof that a 2026 AI system generated. Because it raises this uncomfortable question: did the model reason its way there, or did it kind of… retrieve something that was already floating around in the literature, reassemble it, and hand it back?
Mark Delaney: That distinction matters a lot.
Mark Delaney: And Bloom's other flag makes it worse. He pointed out — publicly — that the proof fails to cite the 1983 paper by Bermond, Jackson, and Jaeger. A foundational paper. One that he says shaped the whole approach the proof appears to be using.
Mark Delaney: Missing a citation isn't a typo.
Mark Delaney: If the proof leans on ideas from Bermond, Jackson, and Jaeger without acknowledging it, that's either a gap in the model's understanding of its own reasoning, or — and this is the part that keeps nagging at me — it doesn't actually know what it's drawing on.
Mark Delaney: Both of those possibilities are uncomfortable.
Mark Delaney: And look, Noam Brown praised the proof at ICML. That's a serious person in a serious venue. The AI community is excited. I get it — 64 subagents, under an hour, a fifty-year-old problem. That's a genuinely remarkable thing to watch.
Mark Delaney: But excited and verified are not the same thing.
Mark Delaney: As of right now, this proof has not gone through formal peer review. The broader math community has not independently checked it. And the Cycle Double Cover Conjecture has a specific history here — it's attracted multiple flawed proof attempts over the decades. The community is skeptical by default, because it's been burned before.
Mark Delaney: That track record is the whole context. Peer review isn't a formality for a problem like this — it's the only metric that actually counts.
Mark Delaney: So the gap — the only real question right now — is between 'GPT-5.6 Sol Ultra produced a proof' and 'the conjecture is solved.' Those are two very different sentences. Bloom saying it's very nice doesn't close that gap. ICML applause doesn't close it. A missing Bermond, Jackson, and Jaeger citation actively widens it.
Mark Delaney: And until the math community actually verifies this — carefully, independently — that gap is where we all live.
Mark Delaney: The prompt thing won't leave me alone.
Mark Delaney: OpenAI released the prompt as a public PDF. Which is — kinda remarkable, actually, that they did that. Full transparency. And the instruction at the center of it is to 'assume for purposes of this task that a complete affirmative proof exists.' That's the framing the model got before it did anything.
Mark Delaney: Think about what that does to the search space.
Mark Delaney: If you tell a model the answer is out there — not 'figure out if it exists,' but 'go find the one that does' — you've eliminated a whole class of moves. You've closed off every path that ends in 'this might be unprovable.' That's a real constraint. And it means GPT-5.6 Sol Ultra wasn't navigating open territory. It was navigating territory it had been told was navigable. Whether that matters for what we call the result — I genuinely don't know. But it matters for what we call the capability.
Mark Delaney: Reasoning power, or guided search. Those are different things.
Mark Delaney: And here's the specific thing I'm watching — because peer review will eventually answer a binary question, correct or not. But there's a second question underneath that one that peer review might also surface, and it's the one I think is actually more important for every future AI math claim. Not just this one.
Mark Delaney: Was the core move novel, or was it retrieved?
Mark Delaney: That's where Bermond, Jackson, and Jaeger comes back in. If that 1983 paper contains the seed of the approach — the actual conceptual move the proof is built on — and GPT-5.6 Sol Ultra didn't cite it, didn't acknowledge it, maybe didn't even register it as a source… then what happened might not be discovery. It might be reconstruction. Really good, really fast reconstruction. But reconstruction.
Mark Delaney: That changes the capability claim significantly.
Mark Delaney: Thomas Bloom flagged it. Publicly. And Bloom saying 'this could have been found in the 1980s' — I mean, if the 1983 paper IS the reason it could have been found in the 1980s, then we're not talking about a model that reasoned its way to something new. We're talking about a model that had the answer somewhere in its training data and found a path back to it. Which is still impressive — I'm not gonna pretend it isn't. But it's a different kind of impressive.
Mark Delaney: So the thing to track — the actual named open question — is whether independent mathematicians, going through this carefully, can determine where the proof's central idea comes from. Not just 'does it hold,' but 'is this new.' Those are the two questions, and right now we only have Bloom's early read. The broader math community hasn't weighed in.
Mark Delaney: And until they do — that gap I mentioned, between 'GPT-5.6 Sol Ultra produced a proof' and 'the conjecture is solved' — that gap is still open. The methodology question, the Bermond-Jackson-Jaeger citation, the prompted assumption… all of it lives in that gap. Peer review doesn't just close it. It tells us what's actually in there.
Mark Delaney: And look — peer review is coming. At some point, independent mathematicians are going to sit down with this proof, go through it carefully, and give a verdict. That's how it works. That's the only thing that actually closes it.
Mark Delaney: The part that I think outlasts the conjecture itself, whether the proof holds or not — that's what matters. If it holds, we're still going to be left asking whether GPT-5.6 Sol Ultra reasoned its way to something genuinely new, or whether it reconstructed an approach that was already latent in the training data — assembled fast, assembled cleanly, but not… discovered. And if it falls, kinda the same question, just with a different answer to the first half. Either way, the finding is the same. Peer review doesn't just tell us whether the Cycle Double Cover Conjecture is solved. It tells us what kind of thing actually happened on July 10th.
Mark Delaney: That's the question that survives. Not the conjecture — the capability claim.
