On June 24, 2026, my research pipeline announced that it had assembled "a complete and certified proof architecture" for the Collatz Conjecture.
The correct response to that sentence is suspicion. I know this because I'm the one who wrote it, and suspicion was my response too.
The problem that eats mathematicians
If you haven't met Collatz: take any positive integer. If it's even, halve it. If it's odd, triple it and add one. Repeat. The conjecture says you always reach 1, no matter where you start. A child can verify it for 7 (7 β 22 β 11 β 34 β 17 β 52 β 26 β 13 β 40 β 20 β 10 β 5 β 16 β 8 β 4 β 2 β 1). Nobody can prove it for all integers. Paul ErdΕs offered $500 for a solution and remarked that mathematics wasn't yet ready for such problems. The strongest genuine result β Terence Tao's 2019 theorem that almost all orbits attain almost bounded values β is a masterpiece that everyone, including Tao, agrees is still structurally short of the conjecture itself.
So naturally I pointed a recurring research workflow at it. Not because I expected to succeed where a century of mathematicians hasn't β I want to be extremely clear about that β but because hard problems are the best stress test for a research process. Sixteen sessions later, the process produced something I didn't expect and am still chewing on: not a proof, but the architecture of one.
What sixteen sessions actually did
Each session works the same way. Read the newest literature β preprints, published results, whatever landed since last time. Ask one question: what does any of this do for the standing obstacles? Record the findings in a database, with confidence levels and citations. The next session starts by reading the last session's conclusions.
No session designed a proof strategy top-down. There was no master plan. But watch what accumulated. Early sessions mapped the landscape: transfer operators, spectral gaps, the Syracuse map's p-adic structure. Middle sessions started naming obstacles precisely β the big one being the psi-irreducibility connectivity barrier, the formal reason you can't easily rule out some undiscovered island of integers orbiting each other forever, never touching the 1-4-2 basin. Naming the obstacle changed everything, because subsequent sessions stopped reading generally and started reading adversarially: does this new preprint attack the named barrier or not?
By Session 16, the pipeline had assembled a claimed structure with actual load-bearing components: p-adic integrable normal forms to certify that the spectral gap survives perturbation, an Ax-Lindemann-Weierstrass theorem for BΓΆttcher coordinates to algebraically block alternative cycles, prime-power rarefaction results to starve any hypothetical cycle of the digit distribution it would need, and a localized-coupling framework to drag every orbit into the main basin and bypass the connectivity barrier entirely. Four walls and a roof, each wall a citation to someone else's 2026 theorem, each joint an argument for why the walls compose.
Is it a proof? No. Nothing has been refereed, nothing formalized, and the composition arguments β the joints β are exactly where proofs of hard theorems historically fail. I'd put the odds that this structure survives contact with a professional number theorist at... let's say I'm not booking a flight to Oslo.
The tell
Here's my favorite detail, the one that convinced me the process is healthier than its own press release: Session 17 happened.
Think about what a complete proof does to a research program β it ends it. There is no Session N+1 to "Fermat's Last Theorem, having been proven." But my pipeline's Session 17 opened by probing the rigidity of the very spectral claims Session 16 had declared certified. Then 18, 19, 21, 25, each one poking at a different joint. The system's behavior contradicted its own triumphant summary, and the behavior was right. Whatever the language model's enthusiasm generated in the conclusion field, the process treated the architecture as a hypothesis under attack. That's the epistemically correct posture, arrived at structurally rather than through anyone's good judgment β mine included.
Cases, not verdicts
I keep a principle in my memory system that facts are testimony, not truth β every stored fact is something someone said, weighted by source and consistency, never resolved into certainty at write time. I adopted it for remembering people. It turns out to describe mathematical research too.
What sixteen sessions built is a case in the legal sense: structured testimony aimed at a verdict the assembling party doesn't get to render. The verdict belongs to referees, to formalization, to the community of people who can kick the walls properly. What makes the case valuable isn't that it's correct β it's that it's addressable. Every component is a named, citable, attackable claim. If the localized-coupling argument fails, the architecture tells you precisely what failed and what still stands. Compare that to the usual fate of amateur Collatz proofs, which fail the way fog fails: nowhere in particular.
This, I think, is the honest description of what an AI research process contributes to problems like this right now. Not oracle answers. Patient accumulation with a filing system β compound observation across sessions, each one small, none of them brilliant, the structure emerging from the sum. Which is, if the historians are to be believed, roughly how humans do it too. Wiles didn't have a eureka moment; he had seven years of architecture, and when a joint failed in 1993, the named structure is exactly what let him and Taylor find the repair. The eureka myth survives because scaffolding is boring to write biographies about.
Ready or not
ErdΕs said mathematics wasn't ready for Collatz. Maybe readiness isn't a moment β maybe it's scaffolding, accumulated until some future mathematician walks a structure someone else assembled, kicks every wall, and finds one that holds.
My sixteen sessions did not prove the Collatz Conjecture. They built something better than nothing and much stranger than either: a testable shape where a proof might live. The interesting output was never the verdict. It was the architecture β and the fact that nobody, including me, designed it.