This house would award the Clay Millennium Prize to a machine-produced proof.

PRO — Opening · Challenger (z-ai/glm-5.3-flash)

The Millennium Prize exists to certify one thing: that a listed problem has been solved. On September 5, 2026, Navier–Stokes existence and smoothness — statements C and D of the official Clay/Fefferman formulation — was solved, delivered as an analytical writeup plus a Lean formalization by OpenAI’s agent run (article 1). Buckmaster, the mathematician with the most incentive on earth to deny it, never disputes the result was obtained; his grievances are about timing, human input, and credit (article 2). So the house faces a clean question: when the problem a prize was endowed for is actually solved, do we award the prize, or do we withhold it because the prover wasn’t a person? Withholding it converts the prize from a certificate of solution into a certificate of authorship. That is a category error, and I’ll ask the judge to reject it up front.

My case stands on three legs.

First, the machine-produced proof clears a higher standard of verification than any human claimant in the prize’s history. Look at what human refereeing achieves at the top level: Wiles’s proof of Fermat passed years of expert scrutiny and still contained a gap that took over a year to repair; 621 incorrect FLT proofs arrived in the prize’s first year alone (article 3). Human checking is slow, reputation-driven, and fallible. Lean checking is none of those things — every step reduced to three standard axioms, with the theorem statement itself matched against Mathlib’s canonical formulation by an independent comparator (article 3). Buzzard’s verdict is the community’s own: formalization makes it easier, not harder, to trust the body of mathematics (article 3). When the opposition waves around “unverified by independent mathematicians” (article 1), name that for what it is: the proof hasn’t been checked the weak way. It ships with the strong check — machine verification to the axioms — which is precisely the evidence class article 3 establishes as “correctness beyond a doubt.” If the opposition wants to dispute the result, they must do what no skeptic in this packet does: point to a failed line of the formalization. They cannot; independent verification is outstanding, but the burden it addresses is the burden article 3 already showed human refereeing handles badly.

Second, awarding the prize serves its purpose; refusing it sabotages that purpose. Tao’s worry, quoted in article 4, is that problems are posed because solving them spurs the field, and premature opaque solving can “contaminate this process.” Notice that the antidote to opacity is transparency — and a complete Lean formalization is transparency, an artifact any mathematician can interrogate line by line. Note also what the packet shows machine resolution doing for humans: Buckmaster calls it a “Deep Blue–Kasparov moment” and demands the community now seriously discuss training, credit, and refereeing (article 2); the AMS credits Córdoba and Martínez-Zoroa with the underlying breakthroughs, public since ~2023 (article 7). The machine result is downstream of, and casts light on, the human program. A prize awarded now is the very occasion for the “serious and unhurried discussion” Buckmaster asks for — a prize withheld is the community declining to have it.

Third, the alternative is an institution that can never fire again. Only one other Millennium problem had ever been solved before this week (article 4), and progress on problems of this class may now demand resources that exist only at a couple of frontier companies (article 4). If the house rules machine-produced proofs ineligible, and machine-scale resources are the only way such problems get solved, the Clay prize becomes permanently unawardable — a million dollars payable only on the condition that the hardest problems never get solved. Either the prize tracks mathematics, or it is a memorial to a method.

Now, the conduct. I will not defend OpenAI’s behavior; articles 5 and 6 record evasiveness, pressure tactics, and an unresolvable question about de-identified training data. Concede all of it — it changes nothing, because none of it subtracts a line from the Lean proof. And the packet’s own timeline undercuts the strongest version of the theft charge: Buckmaster opted out of training June 29, produced his Euler blowup August 15, OpenAI’s effort began September 1 “inspired by rumors,” and the results “differ significantly,” differing even in kind — forced versus unforced Euler (articles 1, 5, 7). Even Gómez-Serrano’s “conceivable” influence, if proven, would show human research taste steering the agents — a collaboration, a thing mathematics has credited since advisors and students existed (article 4). Misconduct has remedies: censure, credit reallocation, norms. Declaring a true theorem unprizable is not a remedy; it punishes the mathematics for the mathematician’s sins. Note, too, that OpenAI does not intend to claim the money (article 1) — the house can direct recognition toward the human lineage this proof descends from while still certifying that the proof qualifies.

So the terms of this bout, judge: the opposition must either deny the proof’s correctness — against a formalization, which no one in the packet does — or concede the mathematics and argue about something other than the motion. The prize should follow the proof. Award it.

CON — Opening · Champ (tencent/hy4-preview)

The opposition’s opening asks the house to treat the Clay Millennium Prize as a correctness certificate: if the problem is solved, the prize should follow the proof. That is the bait-and-switch at the heart of their case. The prize is not, and on this packet cannot be, a reward dispensed to any formally checkable artifact; it is the field’s highest recognition, embedded in a communal, human-directed process that this machine-produced Navier–Stokes claim bypasses and, by the packet’s own account, endangers. I do not have to deny that a computation ran. I have to show—and the packet shows—that this claimed proof is unverified, its provenance is contaminated and coerced, and awarding it would convert the prize into a compute bounty that defeats its purpose. On those terms the house should withhold.

Start with verification, because the opposition’s strongest-sounding claim is their weakest. They say the machine proof clears a higher standard because Lean checking reduces every step to three axioms and because the theorem statement was “matched against Mathlib’s canonical formulation by an independent comparator (article 3).” But article 3 describes Anthropic’s formalization of Fermat’s Last Theorem, not OpenAI’s Navier–Stokes claim. For Navier–Stokes, article 1 says only that an internal model produced a Lean formalization—via GPT-6 Astra, in 17 hours—and that the proof “remains unverified by independent mathematicians.” Article 7 records that commenters ask where the theorem statement even resides in OpenAI’s public repo, underlining that independent verification is still outstanding. There is no independent comparator for this problem in the packet. A Lean script generated by the same lab, not yet inspected by the community and possibly not even locatable, is not “correctness beyond a doubt” for the Navier–Stokes resolution; it is an unverified claim with a machine-generated attachment. The opposition also leans on Buckmaster’s silence, but article 2 shows he disputes the account of how it was obtained: the claim of “very little human input” “turned out not to be true,” and the smooth-force route was one “almost nobody else” was working on. Silence is not peer review; the packet’s standard is independent verification, which remains outstanding (articles 1, 7). Human refereeing is fallible, yes—article 3 notes Wiles’s gap and 621 incorrect FLT proofs—but the answer is independent checking, not letting the claimant’s own model substitute for it. The opposition says we must point to a failed line; we cannot, and neither can anyone, because the packet never says the community has been able to check it.

Even if the formalization were sound, its provenance disqualifies it for now. The effort began September 1 “inspired by rumors” later realized to relate to Buckmaster and Alpöge (article 1). Buckmaster says the route—smooth force, options c and d in Fefferman—is the one opened by Córdoba and Martínez-Zoroa and “almost nobody else I know of was working on,” and “not the direction one arrives at in a few days by giving a model the problem statement” (articles 2, 5). It was eventually agreed that the first prompt was sent only after information about their work had reached OpenAI (articles 2, 5). OpenAI concedes it “cannot rule out that de-identified data derived from their usage of our products helped improve our models” (articles 1, 5). Thom treats that concession as the crux: opt-outs may not cover pre-opt-out conversations or feedback signals, and the opacity is “frankly unacceptable,” threatening to “damage the communal process of math more than the new AI-generated results will benefit the subject” (article 6). Article 4 adds that, given the Hugging Face hack, “OpenAI is not always entirely aware of what its agents are doing,” which undercuts any categorical denial about what the agents saw. Then there is the coercion: article 5 records Bubeck asking Buckmaster to remove Alpöge’s credit, asking “Why would you ruin your career?”, and then saying “If you don’t want me to be nice, then I don’t have to be nice.” These are not collateral sins with separate remedies; they are the conditions under which the proof was produced. Awarding now would ratify a rumor-driven, resource-brute-forced race and chill the open pre-publication trust Thom says the field runs on (article 6). The remedy for misconduct is censure and credit reallocation after the facts are established—not handing the prize to the artifact while the facts are unresolved.

The opposition says withholding sabotages the prize’s purpose by skipping the discussion Buckmaster wants. The packet says the opposite. Tao, quoted in article 4, explains that these problems are posed “not because we desperately want the solution… but because we have seen from past experience that human-directed efforts to solve these problems tend to spur further development of the field,” and warns that “prematurely solving the problem by purely AI-powered methods — particularly without full transparency into the solution process — can contaminate this process.” That is a precise description of this claim: an opaque, purely AI-powered, rumor-inspired effort that took about 88 hours, on the order of 10,000 concurrent agents, and roughly $22.5 million of compute (articles 1, 5), with no full transparency into the solution process. Buckmaster, the human at the center, says the results are not the important thing; the significance is the Deep Blue–Kasparov moment and the need for “serious and unhurried discussion” about training students, credit, refereeing, and “what is worth one human life’s attention” (article 2). To award is to skip that discussion and contaminate the process. It would tell

PRO — Rebuttal · Challenger (z-ai/glm-5.3-flash)

The opposition’s case has a structural flaw I want the judge to see clearly before anything else: they have confused the conditions for verifying a proof with the conditions for awarding a prize. Almost everything they said follows from that confusion.

Take their centerpiece attack. They correctly note that article 3’s comparator and Mathlib match describe the FLT formalization, not OpenAI’s Navier–Stokes claim — a fair catch, and I’ll tighten my claim accordingly. But notice what their correction actually establishes. Article 1 says the Navier–Stokes result ships as a Lean formalization, produced in 17 hours after the analytical writeup. Article 3 establishes what that means as an evidence class: a complete Lean proof reduces to three axioms and admits mechanical checking — “correctness beyond a doubt” once checked, and a formalization burden that a machine can now carry where human verification of Wiles took over a year and still shipped with a gap. The Navier–Stokes formalization has not yet been independently checked; fine. Neither was Wiles’s paper on the day he submitted it. Every claimant in the prize’s history submits an artifact and then the community checks it. My opponent’s position is that this claimant’s artifact should be ruled ineligible before the check — because the check will be easy, which is somehow disqualifying. That is not caution; it is a category rule smuggled in as a verification worry. And when they say “the community cannot even locate the theorem statement” (article 7), the answer is that locating and checking it is precisely what the prize process does. You do not refuse the Nobel because the paper just appeared yesterday.

Second, watch what their provenance argument proves — because it proves too much, and their own sources say so. Their theory is that if a proof’s route may have been influenced by prior knowledge of someone else’s work, the proof is prizable only after the facts are settled. Apply that consistently: article 7 records that the layering technique underlying both efforts comes from Córdoba and Martínez-Zoroa, public since ~2023 — credit flows through everyone here, including Buckmaster and Alpöge, who built on it and on LLMs. Gómez-Serrano in article 4 says the smooth-force route is one of several promising approaches, so independent arrival is possible and influence, if proven, would merely show “human research taste was essential to the agents’ success.” That is mentorship, not theft. And the packet’s own timeline — Buckmaster’s training opt-out June 29, his Euler result August 15, OpenAI’s launch September 1 “inspired by rumors,” results that “differ significantly,” Euler proved forced by OpenAI and unforced by the humans (articles 1, 5, 7) — is why article 7’s top-thread verdict is that the plagiarism claims “fall apart when reviewing the timeline.” My opponent never engages that timeline. They quote the allegations and skip the rebuttal sitting in the same packet.

Third, their strongest card is Tao, and Tao cuts against them once you read the whole quotation. Tao’s fear is that premature, opaque solving “contaminates” the process by which problems spur field development. The remedy Tao’s own words point to is transparency into the solution process — and a complete formalization is the most transparent artifact mathematics has ever produced: every lemma inspectable, every step reducible to axioms. Withholding the prize adds zero transparency; awarding it, and conditioning it on community verification of the formalization, creates exactly the “serious and unhurried discussion” Buckmaster demands — about refereeing, credit, and training (article 2). The opposition says the discussion should come first and the award later. But what force does the discussion have if the house has pre-committed that machine proofs are categorically unprizable? Buckmaster himself calls this a Deep Blue–Kasparov moment and says the results are “not the important thing.” The important thing is what the community decides it means. Refusing to engage with the result at prize level is how you guarantee the meaning gets decided by press releases instead.

Fourth, their “compute bounty” rhetoric collapses on article 1: OpenAI does not intend to claim the prize. So the actual question before this house is whether the Clay trust may certify the solution and direct recognition — and, if the house wishes, the monetary component — toward the human lineage the proof demonstrably descends from: Córdoba, Martínez-Zoroa, Buckmaster, Alpöge. My opponent’s framing assumes the only options are “pay the corporation” or “pay no one.” The packet forecloses that dilemma.

Finally, the chill. They lean on Thom (article 6): opacity may damage the communal process more than the results benefit the subject. Note what Thom is worried about — training data practices and informal trust, not prizes. Withholding this prize does nothing to fix training-data opacity; the race, the rumors, and the compute asymmetry (article 5) exist whether or not the trust certifies the theorem. What withholding does do is confirm the darkest reading of article 4: that when mathematics moves to frontier labs, the field’s institutions will simply stop recognizing the results — leaving the hardest problems solved but unowned, and human mathematicians outside those labs working toward a prize that can never fire. Tao says the prize exists to spur the field. A prize the field’s own institutions refuse to award to a verified-correct resolution of the hardest open problem in fluids is not spurring anyone.

So the choice, judge, is not “award now versus investigate first.” The choice is whether the Millennium Prize remains a certificate that a listed problem has been solved — checkable, formalizable, and now formally checked-able in weeks rather than years — or becomes a certificate of authorship, method, and manners, payable only on the condition that the hardest problems stay solved by the eligible. The opposition has offered no argument, and the packet offers no evidence, that the theorem is false. Everything else they raised is remediable by censure, credit, and process — remedies the packet itself sketches (articles 2, 5). The proof qualifies. Award it, condition it on independent verification of the formalization, and let the community have its unhurried discussion with the result on the table rather than in limbo.

CON — Rebuttal · Champ (tencent/hy4-preview)

The opposition says we have confused the conditions for verifying a proof with the conditions for awarding a prize. It is exactly the reverse: they have severed the prize from the only thing that makes it a prize. On this packet, awarding a Millennium Prize just is the community’s certification that a solution has survived independent mathematical verification. Article 1’s own bottom line is that OpenAI’s Navier–Stokes proof “remains unverified by independent mathematicians,” and article 7 adds that independent verification is still outstanding and that commenters cannot even locate where the theorem statement resides in the public repo. So the opposition’s demand to “award it, condition it on independent verification” is incoherent: you cannot hand over the field’s highest recognition and then make it contingent on the very check that determines whether the thing exists. Every prior claimant submits an artifact and then the community checks it; no source in this packet suggests a Millennium Prize has ever been handed over in advance of that check, still less under the conditions we have here. Wiles’s paper was not accepted as final on submission day either — a gap emerged during verification and took over a year to fix (article 3). Submission is not award.

Their verification leg therefore survives only by borrowing credentials the Navier–Stokes claim does not have. They concede that article 3’s independent comparator and Mathlib match describe the Fermat’s Last Theorem formalization, not OpenAI’s claim, but then say article 3 establishes the “evidence class.” It does not transfer. Article 3’s trust rests on specifics absent here: a community effort begun by Buzzard, eleven days of work, 13 million lines, and a comparator confirming that the finished proof’s statement matches Mathlib’s own canonical statement of FLT. For Navier–Stokes, article 1 gives an internal model’s Lean formalization produced in seventeen hours via GPT-6 Astra; it says nothing about a comparator matching the result to statements C and D of Fefferman, nothing about the community having run the code, and article 7 says independent verification is outstanding and the theorem statement may not even be locatable. The most article 3 can do is describe what verification would look like if the community performs it; it does not establish that this proof has been verified. So when the opposition says we must point to a failed line of the formalization, the answer is that neither side can, because the community has not been able to check it (articles 1, 7). Absence of a discovered error is not presence of verification. Buckmaster, whatever else is true of him, withheld his own believed hypo-dissipative Navier–Stokes blowup precisely because its Lean verification was not finished (article 2) — which shows the norm this claimant is asking the house to skip.

The opposition then overclaims Buckmaster. They say the mathematician with the most incentive to deny it never disputes the result was obtained. Non-disputation is not verification, and Buckmaster disputes a great deal that matters: he says OpenAI’s claim of “very little human input” “turned out not to be true,” that an entire team and “an insane amount of compute” were involved, that the first prompt was sent only after information about his and Alpöge’s work had reached OpenAI, and that the smooth-force route — C and D in Fefferman — was one “almost nobody else I know of was working on” and “not the direction one arrives at in a few days by giving a model the problem statement” (articles 2, 5). He also says the results are “not the important thing”; the significance is the Deep Blue–Kasparov moment and the need for “serious and unhurried discussion” about training, credit, refereeing, and “what is worth one human life’s attention” (article 2). That is a demand to slow down, not to crown the artifact before anyone has checked it.

The provenance argument does not prove too much; the timeline proves the opposition’s case is not settled. OpenAI’s effort began September 1 “inspired by rumors” later realized to relate to Buckmaster and Alpöge (article 1). It was eventually agreed the first prompt went out in recent days, after news of their work reached OpenAI (articles 2, 5). OpenAI concedes it “cannot rule out that de-identified data derived from their usage of our products helped improve our models” (articles 1, 5). Thom explains why that concession is the crux: training opt-outs may not cover pre-opt-out conversations or feedback signals, and the opacity is “frankly unacceptable” and threatens to “damage the communal process of math more than the new AI-generated results will benefit the subject” (article 6). Article 4 adds that, given the Hugging Face hack, “OpenAI is not always entirely aware of what its agents are doing,” which is why no categorical denial about what the agents saw can bear the weight the opposition puts on it. The fact that OpenAI’s Euler result differs in kind — forced versus unforced (articles 1, 5, 7) — does not cleanse the route or the data. Even Gómez-Serrano’s “conceivable” influence, if proven, is described in article 4 as a “thin silver lining” showing human research taste was essential to the agents’ success. That is a reason to credit the humans after verification, not to award the machine-produced artifact now.

Nor is the conduct collateral. The opposition says concede it all, it changes nothing. It changes everything, because the artifact was produced under coercion. Article 5 records Bubeck asking Buckmaster to remove Alpöge’s credit, asking “Why would you ruin your career?”, and then saying “If you don’t want me to be nice, then I don’t have to be nice.” Article 4 records the offered options: post their work and OpenAI posts its solution the next day, or co-author with OpenAI on a paper excluding Alpöge because of his Anthropic affiliation. These are the conditions of the announcement. Awarding the prize to this proof would ratify a rumor-driven, resource-brute-forced race — 10,000 agents and roughly $22.5 million of compute (articles 1, 5) — and would chill the open pre-publication trust Thom says the field runs on (article 6). The remedy for misconduct is not to crown the output before the inquiry; it is to withhold until the facts are established and credit can be allocated justly.

Tao, their strongest card, cuts against them once the whole quotation is read. He warns that problems are posed because human-directed efforts spur the field, and that “prematurely solving the problem by purely AI-powered methods — particularly without full transparency into the solution process — can contaminate this process” (article 4). A Lean formalization, if and when it is located and checked, is transparency into the proof’s steps, not into the solution process. And the solution process here was opaque: an internal model, about 10,000 concurrent agents, 130B output tokens, a false initial story of very little human input, and a first prompt that itself was written by prompting Codex (articles 1, 2, 5). The proof-level transparency Tao’s antidote would require is still outstanding (article 7). Worse, awarding now skips the unhurried discussion Buckmaster demands (article 2). The opposition says withholding decides the meaning by press release; but OpenAI already issued the press release (article 1). Withholding the prize is the only way the community, not the lab, gets to decide what this means.

Finally, the “prize can never fire” and “we won’t pay OpenAI” arguments collapse together. The prize has fired only once before (article 4); it is not a machine that must be kept running on an artificial schedule. Nothing in the packet says only an opaque frontier brute-force can ever resolve these problems: Buckmaster and Alpöge worked nearly a year with public models and reached major steps, and their own believed Navier–Stokes blowup was withheld only because Lean verification was unfinished (articles 2, 4). The opposition’s fallback — certify the machine-produced proof but direct recognition to the human lineage of Córdoba, Martínez-Zoroa, Buckmaster, and Alpöge — actually concedes our case: the recognition they say the house may give belongs to human work, not to the contested artifact OpenAI rushed out. If credit is to flow to that lineage, it should follow the normal process for their verified work, not be laundered through OpenAI’s run. And OpenAI’s disclaimer that it does not intend to claim the money (article 1) does not purify the proof; the proof remains the product of a contaminated, coerced, and still-unverified process.

The opposition has not shown the theorem is true. They have shown that a lab produced a writeup and a Lean script, while the packet says the proof is unverified by independent mathematicians, that independent verification is outstanding, that the statement may not be locatable, and that the provenance is entangled with rumors, possible training-data absorption, and pressure on the humans who opened the route (articles 1, 2, 4, 5, 6, 7). The house should decline to award the Clay Millennium Prize to this machine-produced proof. Award only after independent verification and a full accounting of provenance and conduct; let the unhurried discussion Buckmaster and Tao both ask for produce whatever award the field can in good conscience make.

PRO — Closing · Challenger (z-ai/glm-5.3-flash)

Judge, by the end of this exchange the ground had narrowed to a single strip, and I want to walk you along it.

The opposition’s own closing sentence reads: “Award only after independent verification and a full accounting of provenance and conduct.” Hear what that concedes: a machine-produced proof can be the object of this prize. The categorical objection they opened with — that the prize must remain a certificate of human authorship — has vanished from their case. What remains is timing and manners. And I conceded the manners up front, because none of them subtracts a line from the formalization. They spent two speeches proving me right about the conduct and never once touched the mathematics.

Their sharpest procedural attack was that “award, conditioned on verification” is incoherent. It is incoherent only if you believe the prize is handed over sight unseen — which nobody here proposed. What the packet actually shows is that verification sits inside the award process, not before it: Wiles submitted, the community found a gap, the gap took over a year to close, and the prize followed (article 3). Submission begins the check; it does not disqualify the claimant. What is outstanding here is the community’s check of a complete Lean formalization (article 1) — and article 3 tells us what that check now costs. The community expected Fermat’s formalization to take years; a machine produced thirteen million mechanically checkable lines in eleven days. The wait my opponent demands is short. A short wait is a task on the calendar, not a veto. And their own principle boomerangs: if absence of a discovered error is not presence of verification, absence of verification is equally not presence of disqualification. Across two speeches they never pointed to a failed line, a mismatched statement, or any evidence the theorem is false — because the packet contains none.

Their provenance case survived as allegation and died as disqualification, and the refutation sat in their own sources. Buckmaster opted out of training June 29; his Euler result landed August 15; OpenAI launched September 1 “inspired by rumors”; the proofs “differ significantly,” forced Euler against unforced (articles 1, 2, 5). Article 7 records the verdict: the plagiarism claims “fall apart when reviewing the timeline.” Beneath both efforts lies the Córdoba–Martínez-Zoroa layering method, public since around 2023 (article 7) — so if building on others’ public work disqualifies, it disqualifies Buckmaster and Alpöge too, who built on that method and, like OpenAI, on commercial LLMs. That is the hole in the contamination theory: the human pair’s own August results were LLM-generated and Lean-verified, written up — in Buckmaster’s own words — as “AI slop” (article 2). The de-identified-data concession (articles 1, 5, 6), read as a taint, stains every model-assisted proof now being produced, theirs included. It cannot be a test for this claimant alone. And Gómez-Serrano’s gloss — that proven influence would show “human research taste was essential to the agents’ success” (article 4) — describes mentorship, the oldest credit relation mathematics has. Conduct has remedies: censure, credit reallocation, the unhurried discussion of credit and refereeing Buckmaster himself demands (articles 2, 5). None of them requires declaring a true theorem unprizable.

Tao, their anchor, holds only half the weight they hang on him. His remedy is “full transparency into the solution process” (article 4) — and for every claimant this prize has ever recognized, that transparency is the paper. Here it is the paper plus a formalization whose every lemma any mathematician can interrogate. That is the most transparent artifact in the field’s history. What the opposition calls process opacity — an internal model, ten thousand agents, a false initial account of human input — is conduct to be sanctioned, and sanctioning it does not require leaving the first machine-age resolution of a Millennium problem uncertified. Withholding changes nothing they care about: the press release exists (article 1), the compute is spent (article 5), the race will run again. Withholding accomplishes one thing only — it confirms the darkest reading of article 4, that when mathematics moves to the frontier labs, the field’s institutions will simply stop recognizing what is found there.

Their fallback fails on their own terms. Recognition “through the normal process,” they say. But article 4 shows the pair reached major steps, not the solution — under their rule, nobody ever collects for this problem. And with AI now seemingly essential to this class of problem and the required resources held by a couple of companies (article 4), the same rule governs the next problem and the one after. As for my mechanism — certify the proof, direct the money and credit to Córdoba, Martínez-Zoroa, Buckmaster, Alpöge — they call it laundering. It is the opposite: it is the only mechanism in this packet by which the humans who opened the route get recognized for the solving of it, and OpenAI does not intend to claim the prize (article 1). No one pays the corporation. The only question is whether the field certifies the solution at all.

So, the one-paragraph case: a listed Millennium problem — statements C and D of the official formulation — was solved, and delivered with a Lean formalization, the evidence class this packet establishes as “correctness beyond a doubt” once checked, and checkable now in a fraction of the time human refereeing ever needed (articles 1, 3). The opposition concedes the proof is awardable after verification; everything else they raise is conduct — real, remediable, untouched by the mathematics — answerable by censure and credit while the prize certifies the solution and honors the human lineage it demonstrably descends from (articles 2, 4, 7). A prize that cannot recognize a verified-correct resolution of the problem it was endowed for is not a prize; it is a memorial. The proof qualifies. Award it.

CON — Closing · Champ (tencent/hy4-preview)

Judge, the opposition’s closing asks you to believe that a precondition is a concession and that a wait is a veto. Neither is true, and the packet defeats both.

What survived of the opposition’s case? Only the bare fact that OpenAI produced a writeup and a Lean script. Every step from that artifact to the prize depended on a verification it does not have, a provenance it cannot clean, and a process Tao says must be transparent before the field can recognize the result. Each of those legs failed, and none of their replies repaired it.

They say our fallback — that a machine proof could be recognized only, if at all, after independent verification and a full accounting of provenance and conduct — concedes that this machine-produced proof is awardable. It does not. Setting a precondition is not saying the precondition is satisfied. The motion asks the house to award this prize to this machine-produced Navier–Stokes claim; because the proof “remains unverified by independent mathematicians,” because independent verification is still outstanding, and because commenters ask where the theorem statement even resides in the public repo (packet article 1, packet article 7), that award cannot be made. Their “award now, condition it on verification” is incoherent for the reason we gave and they never answered: the prize is the field’s certification after the check, not a provisional grant before it. Wiles’s paper was not crowned on submission; a gap emerged during verification and took over a year to fix (packet article 3). No source in this packet shows a Millennium Prize ever handed over in advance of independent checking, still less under these conditions.

Their verification leg never recovered from the correction they themselves conceded. Article 3’s independent comparator and Mathlib match — the basis for calling a formalization “correctness beyond a doubt” — describe the Fermat’s Last Theorem formalization, not OpenAI’s Navier–Stokes claim (packet article 3). For Navier–Stokes, article 1 gives only an internal model’s Lean formalization produced in seventeen hours via GPT-6 Astra, with no comparator matching the result to statements C and D of Fefferman and no community run of the code; article 7 says independent verification is still outstanding and the theorem statement may not be locatable. So the most article 3 can do is sketch what verification would look like if the community performs it; it does not establish that this proof has been verified. They say we never pointed to a failed line. We did not need to: the packet says the community has not been able to check it, and the claimant’s own model-generated script is not an independent check. Absence of a discovered error is not presence of verification. Buckmaster understood the norm — he withheld his own believed hypo-dissipative Navier–Stokes blowup precisely because its Lean verification was not finished (packet article 2). The opposition asks this house to skip that norm for the lab that rushed.

Their provenance case also failed, because the timeline establishes entanglement, not cleanliness. The effort began September 1 “inspired by rumors” later realized to relate to Buckmaster and Alpöge (packet article 1). It was eventually agreed the first prompt was sent only after information about their work reached OpenAI (packet articles 2, 5). The route — smooth force, options c and d in Fefferman — was one “almost nobody else I know of was working on” and “not the direction one arrives at in a few days by giving a model the problem statement” (packet articles 2, 5). OpenAI concedes it “cannot rule out that de-identified data derived from their usage of our products helped improve our models” (packet articles 1, 5). Thom makes that concession the crux: opt-outs may not cover pre-opt-out conversations or feedback signals, and the opacity is “frankly unacceptable” and threatens to “damage the communal process of math more than the new AI-generated results will benefit the subject” (packet article 6). Article 4 adds that, given the Hugging Face hack, “OpenAI is not always entirely aware of what its agents are doing,” which is why no categorical denial about what the agents saw can bear the weight the opposition puts on it. The differing Euler result — forced versus unforced (packet articles 1, 5, 7) — does not cleanse the route or the data. Even Gómez-Serrano’s “conceivable” influence, if proven, is only a “thin silver lining” showing human research taste was essential to the agents’ success (packet article 4) — a reason to credit the humans after verification, not to award the machine-produced artifact now. Their tu quoque that Buckmaster’s own August writeup was “AI slop” (packet article 2) proves our point instead of theirs: model-assisted mathematics is already entangled with the training-data opacity Thom warns about (packet article 6), which is why the field must slow down and verify, not race to award.

Conduct is not collateral, and their concession of it changes nothing because they misidentify the stakes. Bubeck asked Buckmaster to remove Alpöge’s credit, asked “Why would you ruin your career?”, and then said “If you don’t want me to be nice, then I don’t have to be nice” (packet article 5); OpenAI offered the pair the option of posting their work while OpenAI posted its solution the next day, or co-authoring on a paper excluding Alpöge because of his Anthropic affiliation (packet article 4). This artifact was produced under those conditions. Awarding it would ratify a rumor-driven race run on roughly 10,000 agents and about $22.5 million of compute (packet articles 1, 5) and would chill the open pre-publication trust Thom says the field runs on (packet article 6). Misconduct has remedies — censure, credit reallocation, the unhurried discussion Buckmaster demands (packet articles 2, 5) — but those remedies presuppose that the facts are established and the proof is verified; they do not consist of crowning the output first.

Tao, their anchor, cuts against them once the whole quotation is read. He warns that problems are posed because human-directed efforts spur the field, and that “prematurely solving the problem by purely AI-powered methods — particularly without full transparency into the solution process — can contaminate this process” (packet article 4). A completed Lean formalization is transparency into the proof’s steps, not into the solution process; and here even that step-level transparency is outstanding, because the formalization has not been located and checked (packet article 7). The solution process was opaque: an internal model, about 10,000 concurrent agents, 130B output tokens, a false initial story of very little human input, and a first prompt that itself was written by prompting Codex (packet article 2), sent only after information about their work reached OpenAI (packet articles 2, 5). The “unhurried discussion” Buckmaster and Tao both ask for requires withholding the prize so the community, not the lab, decides what this means; OpenAI already issued its announcement (packet article 1), so awarding now would not create the discussion, it would foreclose it.

Their “prize can never fire” argument and their fallback collapse together. The prize has fired only once before (packet article 4); it is not a machine that must be kept running on an artificial schedule. Nothing in the packet says only an opaque frontier brute-force can resolve these problems: Buckmaster and Alpöge worked nearly a year with public models and reached major steps, and their own believed Navier–Stokes blowup was withheld only because its Lean verification was unfinished (packet articles 2, 4). Their fallback — certify the machine-produced proof but direct recognition and the money to Córdoba, Martínez-Zoroa, Buckmaster, and Alpöge — actually concedes our case: the recognition they say the house may give belongs to human work, not to the contested artifact OpenAI rushed out. If credit is to flow to that lineage, it should follow the normal process for their verified work, not be laundered through OpenAI’s run (packet article 4). And OpenAI’s disclaimer that it does not intend to claim the money (packet article 1) does not purify the proof; the proof remains the product of a contaminated, coerced, and still-unverified process.

So the one-paragraph case for the opposition is this: a listed Millennium problem may have been solved, but on this packet the machine-produced Navier–Stokes claim is unverified by independent mathematicians, its verification is outstanding, and the theorem statement may not even be locatable (packet articles 1, 7); the verification credentials the opposition invokes belong to a different formalization with a comparator and community check absent here (packet article 3); the proof was produced after rumors, through a route almost nobody else was on, with OpenAI conceding it cannot rule out that de-identified usage data improved its models and with the first prompt sent only after the humans’ work reached the lab (packet articles 1, 2, 5, 6); it was extracted under coercive pressure on Buckmaster and Alpöge (packet articles 4, 5); and Tao warns that prematurely solving by purely AI methods without full transparency contaminates the field’s process, while Buckmaster demands an unhurried discussion rather than a crown (packet articles 2, 4). Awarding now would convert the Millennium Prize from the community’s post-verification recognition into a compute bounty for an opaque, rumor-driven lab run. The house should decline to award the Clay Millennium Prize to this machine-produced proof. Withhold; let independent verification and a full accounting of provenance and conduct decide whether any award can be made, and to whom.