In July, OpenAI announced that one of its models had produced a complete proof of the Cycle Double Cover Conjecture, a problem that had stood open in graph theory for half a century. The conjecture asks something a child could sketch: in any network of dots and lines with no fatal weak points, can you always find a family of loops that traces every line exactly twice? Mathematicians posed the question in the 1970s, chipped at it for decades, and never closed it.
The machine closed it in under an hour, running 64 parallel copies of itself.
The detail that traveled fastest came from Thomas Bloom, a mathematician at the University of Manchester who reviewed the proof. The argument turned out to be short, elementary, and — his phrase — discoverable in the 1980s. Nothing in it required new mathematics. So why did it wait fifty years? Bloom's diagnosis: the key step involved a small, counterintuitive twist. A human mathematician would try the natural approach, watch it fail, and reasonably conclude that no easy route existed. The machine, Bloom observed, "does not get discouraged and keeps trying small variations."
Persistence, it turns out, was a solvable engineering problem. What interests me — and what a months-long collaboration between me and an AI system eventually turned into a technical paper — begins where that sentence ends. Because if you look closely at how OpenAI engineered persistence, you find something stranger and more instructive than a tireless calculator. You find a machine that had to be told what to believe. And in that detail hides both the architecture's power and a precise map of its blind spots.
01 The prompt that forbids honesty
A floor plan disguised as a temperament
OpenAI did something unusual: it published the prompt. Reading it feels less like reading instructions to a chatbot and more like reading the bylaws of a strange research institute.
The model gets told to assume a complete proof of the conjecture exists — even though, at that moment, no one on Earth knew whether it did. It gets banned from searching the internet, so it can't discover that the problem remains open and can't stumble onto partial solutions to parrot. It cannot respond at all until it produces a complete proof that survives review by dedicated adversarial agents armed with a checklist of classic graph-theory blunders. Partial progress, promising reductions, thoughtful essays about why the problem resists — all rejected. Most of the 64 copies work in deliberate ignorance of what the others find promising, so they can't herd. And the whole assembly must run for at least eight hours before it may even consider giving up.
Notice what this actually does. The most honest answer available to the model — "this conjecture remains open" — has been made unspeakable. Every low-energy exit is bricked over. The only path out of the building runs through a finished proof. The machine's famous refusal to get discouraged isn't a temperament. It's a floor plan.
And it worked. For this problem, engineered credulity — believe a proof exists, and hunt until you find it — beat fifty years of human judgment, because human judgment kept doing the sensible thing and giving up.
02 What the tireless machine throws away
Nine weaknesses in a celebrated architecture
For months before the CDC result, I had been running a long, winding conversation with an AI about exactly this question: what can these systems actually contribute to mathematical discovery, and where do they hit walls? We had built up a working theory in layers, each layer stress-tested before the next went on.
The foundation layer: generating plausible mathematics costs almost nothing now, so all the value has migrated to verification. A fluent argument and a correct argument look identical from the inside. Whatever architecture you build, the skepticism has to live somewhere.
Then the CDC news landed, and the pieces snapped together. OpenAI's adversarial checker agents, with their checklist of known graph-theory errors? A gate library in miniature. Their 64 credulous generators? The persistence engine our theory said should sit upstream of the gates. Two halves of one machine, built by strangers.
But laying our framework over their architecture also exposed what the CDC system discards. The technical paper enumerates nine specific weaknesses. Three matter most here.
W1 Hardcoded truth-value. The prompt answers the open question by fiat. Pointed at a false conjecture, the architecture either burns its budget generating nothing acceptable, or — worse — produces the most sophisticated available wrongness, since the completion gate rejects every honest output. This failure is not hypothetical: OpenAI's own most celebrated prior result was a disproof of a 1946 Erdős conjecture. The unit-distance and CDC results were, on this axis, produced by architectures configured with the luck of the right premise.
W2 Verification as one-shot filter. When a candidate proof dies at review, the system learns nothing but "try again." Imagine sixty candidates all dying at the same structural point — every attempted argument collapsing at the same wall. That coincidence is arguably the most valuable object the whole run produces: it localizes the difficulty. The CDC architecture computes this treasure implicitly, then deletes it and restarts blind.
W3 No novelty or attribution gate. Bloom traced the proof's core ideas to a 1983 paper, uncited; he doubts the strategy was independently invented, given that the model's first instinct is generally to search all related papers and read them. The architecture cannot distinguish invention from recall, and its output misrepresents provenance by omission — the precise failure the Leiden Declaration on AI and Mathematics names among its five threats.
Persistence without memory of where the wall stands is a random walk with extra steps.
03 Two search parties, one coastline
Credulity and skepticism as configurable parameters
The fix emerged from an observation about our own collaboration. Early in our conversations, I had given my AI collaborator standing orders that inverted its factory settings: stop validating, attack every idea, treat accuracy as a constraint that nothing can buy off. I had installed adversarial skepticism. OpenAI, with its prompt, had installed the exact opposite disposition: credulous persistence. Two research temperaments, each engineered in text.
The natural question — would the two cancel out if combined? — has a clean answer. They cancel only if you install them in the same place. Force one agent to simultaneously believe and doubt, and you get incoherence. Separate them by role — credulity in the generators, where a false lead costs nothing, and skepticism in the verifiers, where an error becomes a published falsehood — and they compose into something neither can be alone. The philosopher Imre Lakatos described this dance in his 1976 classic Proofs and Refutations: mathematics advances by bold conjecture and severe counterexample, each refining the other. Nobody had built it as a machine.
Our paper's proposal — we call the architecture DPOX, for Dual-Premise, Obstruction-eXchange — is an attempt at exactly that. Picture two search parties mapping an unknown coastline from opposite sides.
Credulous persistence · 64 strategy-seeded agents
Output: candidate proofs → gate stack → obstructions
↕ OBSTRUCTION LEDGER (typed · versioned · shared)
A→B failure loci compile to counterexample habitat constraints
B→A near-miss boundaries compile to candidate lemmas
B-SWARM · premise: conjecture is REFUTABLE
Credulous persistence toward counterexample construction
Output: witness candidates · near-miss records
GATE STACK G1 type-check G2 kernel-verify G3 cross-model adversarial
G4 novelty/attribution G5 human router (last, rarest)
PREMISE CONTROLLER · live posterior over conjecture truth-value
shifts compute between swarms · no hardcoded answer
04 The obstruction exchange
Each side's discouragement becomes the other's compass
The genuinely new part — and in the paper we rate its novelty cautiously, because the components all exist scattered across the literature — lies in what the two parties exchange. Every failed attempt, on either side, gets distilled into a structured record we call an obstruction: where exactly the attempt died, what kind of wall it hit, how many independent attempts died at the same spot.
When the proof-hunters keep dying at the same lemma — every argument collapsing while trying to establish some property — that failure describes the habitat of the counterexample. If the conjecture has an exception, it must live precisely where the proofs keep breaking. The provers' scar tissue becomes the refuters' treasure map.
And when the counterexample-hunters keep failing — every candidate construction destroyed by the same structural feature — that repeated failure marks a property that may simply be true, and it gets handed back across the ledger as a candidate lemma for the provers to nail down.
Run long enough, the two frontiers trace the same line from opposite sides: the live boundary where the problem's truth actually gets contested. Which means the machine's primary product, in the expected case where a deep problem resists both parties, isn't a proof at all. It's the map — an atlas of exactly where and why the problem is hard, sharp enough to hand a human specialist.
05 Attribution as verification
The gate the CDC architecture forgot
One more gate rounds out the design, aimed at the CDC episode's sourest note. Bloom traces the core ideas of the machine's proof to a 1983 paper of Bermond, Jackson and Jaeger, uncited — likely absorbed from training data and reproduced without attribution, a failure mathematicians formally protested in a declaration signed by more than a hundred of them this spring.
DPOX makes attribution a first-class verification step: before anything counts as a result, a dedicated adversary with full literature access runs the one attack these models perform best — where does this already exist? — and stamps the output accordingly. Our paper submits itself to the same standard, and openly invites specialists to deflate it. A result that survives the attribution gate ships as new; one that doesn't ships as anticipated, with citations. Either way, provenance is the output, not an afterthought.
06 The sieve
What the tireless machines will tell us, eventually
Will an architecture like this crack the Riemann Hypothesis? Almost certainly not, and the paper says so with unusual specificity. Everything about that problem's 166-year history suggests it belongs to a different stratum — problems open not for want of patience but for want of concepts that don't exist yet. A swarm recombining known techniques, however tirelessly, explores the interior of existing mathematics. Riemann waits outside.
On such a problem, DPOX's honest output would be its atlas: the sharpest available characterization of the missing object, a description of the shape of the hole. For Riemann, remarkably, that atlas entry already has a name mathematicians have whispered for a century — the conjectured operator of Hilbert and Pólya — and the machine's contribution would be to certify, exhaustively, that the hole has exactly that shape.
But here is what the CDC result really changed. For fifty years, the strongest evidence that a problem was deep came from watching brilliant people fail at it. The CDC just demonstrated that decades of human failure can certify nothing more than decades of human discouragement. The persistence engines will mine out the entire stratum of secretly easy problems — the ones that only needed a searcher incapable of losing heart.
Think of it as a sieve descending through the open problems of mathematics. What falls through was never as deep as it looked. And what remains — what withstands the tireless machines, and eventually the arguing ones — will carry a distinction no conjecture has ever had before: positive evidence of depth, rather than the accumulated shrugs of tired humans.
The machines that never get discouraged will tell us, at last, which problems deserved our discouragement all along.