Question ledger

Living map for this package. Primaries are decision-relevant. Secondaries expose assumptions. Rabbit holes are opened only when they change the answer. Status as of 2026-09-08.

Confirmed vs unverified is stated on P1. Clay has not awarded a prize. Independent community acceptance has not had time to form.

Primary 1 — What did OpenAI actually claim, and on which official statement?

Direct answer: On 8 September 2026 OpenAI published an analytical proof and a Lean formalization that three-dimensional incompressible Navier–Stokes, with positive viscosity, can develop unbounded velocity in finite time from rest, under a smooth external force, while kinetic energy stays bounded. The company maps this to Fefferman’s official alternatives (C) on \(\mathbb{R}^3\) and (D) on the torus \(\mathbb{R}^3/\mathbb{Z}^3\). It does not claim unforced global regularity or unforced blowup (A or B).

Secondaries
1. Does Fefferman’s (C)/(D) explicitly allow a smooth force? Yes. (A) and (B) set \(f \equiv 0\); (C) and (D) exist precisely so a breakdown example may include a smooth force obeying the decay/periodicity conditions.
2. Is “smooth force” a loophole? It is the written prize target, not a later invention. Many specialists still treat unforced blowup as the physically central question. Both facts can be true.
3. Does the construction start from rest? OpenAI’s Theorem 1.1 uses zero initial velocity and a compactly supported smooth force.
4. Is energy bounded? Yes: uniformly bounded kinetic energy through the singular time is part of the claimed theorem, matching Fefferman’s physically reasonable energy condition on \(\mathbb{R}^3\).
5. Periodic vs whole space? Compact support of the constructed fields is used to get the torus statement (Corollary 10.6 in the manuscript).
6. Has Clay accepted it? No. Clay’s rules require publication in a qualifying outlet, two years, and general acceptance. OpenAI says it does not intend to claim the prize. Martin Bridson told New Scientist evaluation will be unhurried.
7. Is Wikipedia’s “disproved” language premature? Yes as of the announcement day. A lab claim plus a Lean repo is not Clay acceptance.
8. What remains open even if C/D hold? Unforced 3D Navier–Stokes (A/B), and whether a physically occurring unforced flow can blow up.

Status: answered from primary sources (OpenAI post, Fefferman PDF, OpenAI PDF Theorem 1.1, Clay prize rules).

Primary 2 — What is the Navier–Stokes existence and smoothness problem, in plain language?

Direct answer: The equations treat a fluid as a continuous medium and apply Newton’s second law to every tiny parcel. The Millennium question is whether a perfectly smooth three-dimensional incompressible flow, with viscosity, must stay smooth forever, or whether the continuum description can produce infinite speed in finite time.

Secondaries
1. What does “incompressible” mean here? Density is treated as constant; volume is conserved; \(\nabla \cdot u = 0\).
2. What is viscosity doing? It is the internal friction that damps shear. The puzzle is whether damping always wins against nonlinear stretching in 3D.
3. Why 3D and not 2D? In two dimensions the analogues of (A)/(B) have long been known (Ladyzhenskaya). The stretching mechanism that makes 3D hard is absent in 2D.
4. What is a singularity / blowup? Velocity (or vorticity) becoming unbounded as time approaches a finite T.
5. What are weak solutions? Leray (1934) constructed global solutions in a weaker sense. Smoothness of those solutions is the open regularity question.
6. How is this different from “solving Navier–Stokes” in engineering? Engineers compute approximate fields on a mesh. The prize asks a universal yes/no about the continuum PDE.
7. Why did Clay include it? Fefferman: fluids are important and our understanding of the equations is primitive; standard PDE methods look inadequate.

Status: answered.

Primary 3 — Why does this matter in mathematics?

Direct answer: It is a flagship supercritical PDE. A resolution, in either direction, is a statement about whether the best-known continuum model of fluids is globally well-posed, and about whether new analytic ideas (or machine-checked constructions) can settle questions that estimates have not.

Secondaries
1. What does supercritical mean? The natural energy bound is too weak to control the scaling-critical regularity (\(\dot{H}^{1/2}\)). Energy does not close the estimate.
2. How does this sit among Millennium problems? Poincaré was solved (Perelman). Navier–Stokes would be the second if Clay later accepts a solution. P vs NP, Riemann, Yang–Mills, Hodge, BSD remain.
3. Why is a counterexample as important as a positive proof? Clay wrote the problem as four alternatives. Either global smoothness or an admissible breakdown counts.
4. What prior partial results framed the field? Local existence; small-data global existence; Caffarelli–Kohn–Nirenberg partial regularity; Escauriaza–Seregin–Šverák \(L^\infty_t L^3_x\) regularity; Tao’s averaged-equation blowup; convex integration nonuniqueness (Buckmaster–Vicol); Albritton–Brué–Colombo nonunique Leray–Hopf solutions with a singular-at-\(t=0\) force.
5. Does a forced blowup “explain turbulence”? No. Turbulence is a statistical, multi-scale phenomenon. Blowup is a pointwise continuum failure. Related, not identical.
6. What would a good explanation of blowup have to show? That the nonlinear self-stretching of the flow, not an infinite hand-imposed force, produces the unbounded velocity, while the residual force stays smooth.
7. Why Lean? Machine-checked formalization changes the social epistemology of long proofs: the kernel checks inferences even if humans cannot read 165 pages quickly.

Status: answered.

Primary 4 — Why do we care outside mathematics? What are the practical applications?

Direct answer: The equations already run aircraft, weather, blood-flow, and combustion codes. A Clay-style theorem does not replace those codes tomorrow. We care because the theorem is about whether the continuum model can fail, which is the foundation those codes quietly assume, and because the method (agent swarms + Lean) may transfer to other hard analysis.

Secondaries
1. Will wings, pipes, or climate models be redesigned next week? No. David Silvester told New Scientist a working result is a “mathematical nicety” for applications; CFD already displaced most wind-tunnel use.
2. Then why do engineers still care? Because every CFD code assumes the continuum PDE is a faithful model down to the grid. A proven continuum blowup is a license to ask where the molecular description must take over.
3. Weather and climate? Forecast models discretize related equations. Better analytic control of regularity would, in principle, bound error growth. That is a long research program, not a patch.
4. Medicine? Blood is not a perfect Newtonian incompressible fluid. The prize statement is about the idealised PDE, not blood rheology.
5. Energy / drag? Drag reduction is an empirical and computational industry. The prize does not hand over a 1% drag formula.
6. What would change practice? New a-posteriori blowup diagnostics; new caution about under-resolved 3D simulations claiming “the PDE did this”; possible hybrid continuum–particle switch criteria — all still research.
7. Global problems linkage (from problems.md): Extreme Weather, Climate, Air Pollution, Access to Energy, Child/Maternal Health (blood flow) sit downstream of fluids. The link is real and indirect.
8. Economic scale? Public figures of order \(10^{18}\) CPU-hours/year on fluids and an ~$8B CFD software market (2025, vendor-site class sources) show the installed base. They do not measure the prize’s immediate ROI.

Status: answered, with the Silvester counter-argument integrated.

Primary 5 — How did OpenAI produce the proof?

Direct answer: An unreleased internal model “significantly more capable than GPT-6 Astra,” wrapped in coordinating agents (on the order of 10,000 concurrent for the Navier–Stokes group), with tools (cached web, code). Agents were split across Clay variants A–D. They first resolved unforced 3D Euler blowup (~100 agents, ~50 hours), then concentrated on Navier–Stokes. Result claimed Saturday 5 September 2026, ~88 hours after launch; Lean via GPT-6 Astra took ~17 more hours. ~2.7 million messages and ~130 billion output tokens on the Navier–Stokes effort; ~300 billion output tokens across all attempted problems that week.

Secondaries
1. Is the model named? No. Only “internal,” “training ongoing since 28 August 2026,” “more capable than GPT-6 Astra.”
2. Cost? Press conference figures conflict in secondary reporting: “millions,” “about $15 million if a customer ran it,” “300 billion tokens / $22.5 million at Astra rates.” Treat cost as order-of-magnitude millions, not a single audited number.
3. Human role? OpenAI describes prompting variants, shifting agents after Euler, cross-pollinating groups with Codex, updating to a further-trained checkpoint. “Very little human input” was contested in Buckmaster’s account of the 6 September calls.
4. Safeguards? OpenAI says frontier evaluation isolation and monitoring stayed in place.
5. Why Euler first? Viscosity-off is a standard stepping stone. Unforced Euler blowup is independently major (not a Clay prize problem).
6. Is Lean the proof or a check? Both: an English/analytical manuscript (~165 pages) plus a Lean 4 project (openai/NavierStokesAndEuler, Lean 4.34.0-rc2, Mathlib). Comparator challenges are provided for independent checking.
7. Could agents have searched user Codex sessions? OpenAI: no specific user data accessed. Caveat they cannot rule out de-identified training signal. Mark Chen denied searching user data.
8. Reproducibility for a third party? Not at the same scale without the unreleased model and the reported compute. The certificate (Lean) is the public check, not a cheap re-run of the search.

Status: answered from OpenAI’s own post; cost figures remain multi-sourced and slightly inconsistent.

Primary 6 — What is the actual blowup mechanism?

Direct answer: A self-similar contracting vortex: fluid spirals inward and stretches axially (“spaghetti”). Radial width shrinks faster than axial length. Speeds grow while the core volume shrinks fast enough that kinetic energy stays finite. Oscillatory pulses in an annulus cancel the singular part of the momentum residual so the external force remains smooth even as velocity blows up.

Secondaries
1. Why must terms “get big yet cancel”? If you define force as the PDE residual, any flow “solves” Navier–Stokes. The hard constraint is that this residual stay smooth (and compactly supported / rapidly decaying).
2. What is the inner core doing? Axisymmetric leading profile: inward spiral, axial outflow away from a dividing layer near \(z=0\), pressure dropping toward the axis.
3. Why slightly break reflection symmetry? Exact midplane symmetry would kill needed shear near \(z=0\); a small axial bias keeps amplification alive.
4. What are the pulses? Spatially oscillatory corrections whose nonlinear momentum fluxes cancel the annulus imbalance — in the lineage of oscillatory stress realization (Daneri–Székelyhidi and related Euler constructions).
5. Is this “convex integration”? Related toolbox (oscillations realizing a stress), not a slogan for the whole proof. The manuscript cites that lineage and also wave-dynamics / centrifugal-instability precedents.
6. Forced vs unforced Euler: OpenAI’s Euler result is unforced compactly supported blowup on \(\mathbb{R}^3\). Buckmaster–Alpöge’s public Euler result is forced. Different theorems.
7. Does viscosity survive in the construction? Yes: the theorem is for every \(\nu > 0\). Angular Reynolds number of the core is claimed to diverge while radial Reynolds number stays bounded.

Status: answered at the level of the published physical description (OpenAI PDF §§1–2). Line-by-line analytic verification is unresolved (community has had hours, not years).

Primary 7 — What does this unlock if the proof holds?

Direct answer: It unlocks a negative well-posedness theorem for the official forced problem; a certified example that viscosity need not prevent continuum blowup under smooth forcing; a public Lean artifact others can typecheck; and a demonstration that multi-agent search plus a proof assistant can finish a Millennium-scale analysis construction. It does not unlock a turbulence theory, a new aircraft, or unforced regularity.

Secondaries
1. For PDE theory? A new constructed singular solution class, and pressure on the unforced problem.
2. For mathematical physics? Evidence that the continuum hypothesis in the PDE can fail even with viscosity and smooth forcing — a precise, not metaphorical, breakdown.
3. For AI research? A capability milestone: original research-level analysis, not contest math, with formal verification. OpenAI frames it as evidence they are in a “next period” of AI progress.
4. For scientific method? Tao’s 2026 warning: answers arriving faster than understanding. This event is the test case.
5. For other Millennium problems? OpenAI says agents were launched at all remaining prize problems; only Navier–Stokes (plus Euler) is claimed resolved. No public Riemann/Yang–Mills/P vs NP claim in the 8 September post.
6. For labs vs academia? Compute-as-priority: once a route is rumored, 10,000 agents can outrun a two-person collaboration. That is a social unlock, not a theorem.
7. For Lean culture? A large fluids formalization in Mathlib’s orbit, plus Comparator challenges.

Status: answered as conditional on the proof surviving scrutiny.

Primary 8 — What did Buckmaster and Alpöge do, and what is the credit dispute?

Direct answer: On 8 September 2026 they posted finite-time blowup with smooth forcing for incompressible porous media, Boussinesq, and 3D Euler, with Lean for the released pieces, following the Córdoba–Martínez-Zoroa forcing program, with heavy LLM help. They did not claim full Navier–Stokes. Buckmaster alleges OpenAI learned of their route, raced it, and mishandled credit/coauthorship (including pressure around Alpöge’s Anthropic affiliation and questions about Codex session data). OpenAI says it started 1 September from rumors, did not see their work until public release, offered a joint announcement under a mistaken belief they had NS, and that even the Euler theorems differ (forced vs unforced).

Secondaries
1. Is the Córdoba–Martínez-Zoroa route obscure? Buckmaster says almost nobody else was on smooth-force Clay C/D. Córdoba told Scientific American he would be surprised if NS were done that way so fast.
2. Did OpenAI offer Buckmaster sole authorship? That is Buckmaster’s allegation about 6 September calls with Sébastien Bubeck. Not independently documented beyond his statement.
3. Codex training? OpenAI cannot rule out de-identified product data helping models; denies looking up specific user data.
4. Tao’s view of Buckmaster–Alpöge? He called it a remarkable achievement and wrote that nothing in principle blocked extension to Navier–Stokes given enough compute and AI — a sentence that reads as prophecy after OpenAI’s post.
5. Anandkumar et al.? A separate 7 September Euler stable-singularity claim via PINNs; Tao flagged it as independent and more “mainstream” numerically. Not the OpenAI construction.
6. DeepMind 2025? Earlier ML discovery of unstable singularities in IPM/Boussinesq families — a prior AI-fluids milestone, not Clay NS.
7. Hypo-dissipative NS? Buckmaster says they believe they have it but did not release pending Lean.

Status: facts of the public papers vs allegations of process — keep separate. Process claims are contested.

Primary 9 — Is the Clay prize actually won?

Direct answer: Not today. Even a correct C/D proof is only a Proposed Solution under Clay rules after qualifying publication, two years, and general acceptance. OpenAI is not claiming the million dollars. Bridson: evaluation will be rigorous and slow.

Secondaries
1. Qualifying outlet? A company PDF and GitHub repo are not automatically a qualifying journal.
2. Could Clay reject C/D as not the “spirit” of the problem? Rules say completeness is CMI’s sole discretion, and NS may be resolved in either direction by the standard procedure. Spirit arguments are community politics, not the PDF.
3. Perelman precedent? Poincaré was solved by a human, prize declined. A corporate AI author is unprecedented for Clay.
4. Who would be the “solver”? Agents, OpenAI-the-author, named researchers, or prior human route-finders? Unresolved socially.
5. Two-year clock? Has not started in any official sense.
6. Fake/crank NS proofs? Decades of them. Lean raises the floor but does not replace conceptual refereeing (wrong theorem, correctly formalized, is still wrong for the intended problem).
7. navier-stokes.org “still open” pages dated March–July 2026 are stale relative to 8 September claims; they are not a live Clay ruling.

Status: answered.

Primary 10 — EmTech taxonomy: LAC, LTC, PTC, milestone?

Direct answer: The underlying EmTech is Artificial Intelligence (with Computing). The new capability is machine-checked research-level theorem construction. GPT-6 Astra is a PTC used for Lean, not the search model. The unreleased internal model is a distinct PTC (unnamed). The 5 September NS resolution is a Milestone on that capability. The LAC is AI-native mathematical discovery. Forced NS blowup is a PAC (a specific constructed solution), not a product.

Secondaries
1. Is this a Life-Altering Capability? Not by itself. It is a milestone toward AI systems that can close century-old analysis problems. Life-altering would be reliable, cheap, general scientific discovery — still unproven.
2. Weak vs strong convergence? Strong-ish: AI × formal methods (Lean) × large-scale agent orchestration. Fluids math is the target, not the second EmTech.
3. Trend class? Landmark solved problems: stepwise. Token/agent scale: too data-poor to claim a doubling time from one event. AI math contest scores over 2024–2026 look closer to exponential capability, but this package will not invent a doubling time.
4. Next milestone? Unforced 3D NS (A/B); independent Lean rebuild from scratch; Clay-process start; a non-fluids Millennium problem.
5. Bottleneck? Human understanding bandwidth (Tao); credit/IP norms; model access; remaining supercritical estimates for \(f=0\).

Status: answered.

Primary 11 — What would refute or weaken the headline?

Direct answer: A gap in the Lean build; a mismatch between the Lean theorem and Fefferman (C)/(D); a smooth-force that violates decay/compact-support conditions; an error in the cancellation construction; or Clay later stating that C/D-with-this-force is not a complete solution.

Secondaries
1. Early technical scrutiny? Secondary report that Gonzalo Cao-Labora requested scrutiny (not a completed refutation). Track, do not treat as a disproof.
2. “Force is cheating”? Invalid as a rules objection; valid as a physics interest objection.
3. Comparator sorry in challenge files? Challenge statements can contain sorry as goals. Do not confuse challenge templates with the filled proof.
4. 165-page unreadability? A social risk, not a refutation. Lean is the intended mitigator.
5. Training-data contamination? Would stain process, not automatically the mathematics, if the certificate checks.

Status: open as a watchlist.

Rabbit holes

Hole	Why opened	Closure
Forced vs unforced as “the real problem”	Changes what “solved” means	Keep both: rules vs spirit. Integrated in 01, 04, 08.
Tao “answers without understanding”	Changes what the event unlocks	Integrated in 06, 09.
Codex / user-data allegation	Changes trust in AI-for-science	Integrated in 07; unresolved factually.
PINN Euler (Anandkumar)	Independent 7–8 Sep fluids AI	Noted as parallel, not OpenAI’s proof.
DeepMind 2025 unstable singularities	Prior AI fluids	Historical, not Clay NS.
CFD market numbers	Practical “why care”	Cited as scale, not prize ROI.
All-In / Moonshots on NS	AI-topic add-on	Shows recorded before 8 Sep; Fermat/Astra signal only.

Unresolved

1. Independent verification of the 165-page analysis and the Lean build by non-OpenAI groups.
2. Whether de-identified Codex/product data influenced the internal model.
3. Clay’s eventual completeness judgment on C/D.
4. Unforced 3D Navier–Stokes (A/B).
5. Exact internal model identity and true fully-loaded cost.
6. Whether community will treat forced blowup as conceptually sufficient.

Stabilization note

No new primary added after the source round that included Fefferman’s PDF, OpenAI’s 8 September post, the NS manuscript front matter, the GitHub README, Clay prize rules, Buckmaster coverage, Tao, Silvester, and the podcast recency check. Contrarian positions (Silvester nicety; “force is not the problem people meant”; credit theft; still-open A/B; Clay delay) are in the body, not a footnote.