2026-10-01
OpenAI's Lean-checked Navier-Stokes blow-up is mathematically clean. Physically, water stops being a continuum at 0.67 nm and air at 45 nm; liquids cavitate much earlier.
On 8 September 2026 OpenAI said an internal multi-agent system had produced Lean 4 proofs of finite-time blow-up for the forced 3D Navier-Stokes equations (Millennium Prize alternatives C and D) and for unforced Euler. The kernel check is clean: zero sorrys, zero custom axioms, no weakened norms. Figures such as "10,000 agents" and "88 hours" appeared in secondary coverage, not in either OpenAI paper, and this audit does not treat them as fact.
The Clay problems ask precise questions about a continuum incompressible model on R3. They were never claims about how water behaves. Fluid dynamicists who read an earlier public draft said the same. The remaining job is quantitative: how far this constructed solution travels from the regime Navier-Stokes is used for, and which constitutive assumption fails first.
Xavier Callens at Socrate AI Lab offers a physical reading, not a hunt for Lean bugs. The collapsing core keeps a radial Reynolds number of order one, so its scales are diffusive: radius sqrt(nu t), speed sqrt(nu/t), almost independent of the initial vortex size or the choice of units. Compressibility, rarefaction and viscous heating then become order-one at one length, l = nu/cs. Liquids hit cavitation earlier still.
Nobody, including this project, has simulated the 166-page OpenAI construction. Section 9 uses a surrogate: a manufactured Gaussian core on the same diffusive scaling, forced by the incompressible Navier-Stokes residual (time-reversed diffusion, which is anti-diffusion). A 3D pseudo-spectral solver hits the Taylor-Green Re=1600 dissipation peak at t=9.14 against a published 9.0. The bundle also runs 1D compressible Navier-Stokes-Fourier, a discrete-velocity BGK solver, and Lennard-Jones molecular dynamics with no viscosity law and no equation of state put in. On the formal side, Lean proofs on OpenAI's own definitions show that any object with their candidate properties exceeds every velocity-gradient bound arbitrarily close to the singular time: 85 declarations in ten files, three standard axioms only.
The continuum dies at a material-specific scale.
| Fluid (300 K) | l = nu/cs | t = nu/cs^2 | max vorticity |
| Water | 0.67 nm | 0.44 ps | 2.2e12 /s |
| Air | 45 nm | 0.13 ns | 7.5e9 /s |
In ambient-pressure water, cavitation arrives first. A Bernoulli pressure deficit 0.5 rho u^2 hits vapour pressure near u 14 m/s (Mach 0.01), about 5 ns before blow-up, in a core still 70 nm wide: three decades earlier in time. A Gaussian core moves the threshold to 7.6 m/s, 130 nm, 17 ns. Once a vapour cavity opens, a hollow vortex caps swirl near 14 m/s.
Global kinetic energy stays bounded as the Prize requires (exponent +0.485, tending to 0). Local energy density tau^{-1.010} and the dissipation rate tau^{-0.515} both diverge. The moment-matching matrix looks like kappa 1e28; after non-dimensionalization kappa(B) 4.11e5, independent of radius. The acoustic-radiation check is Ma^5, the Mach criterion again. The force is the smooth remainder of a co-designed (u, f) pair, with no independent physical origin.
A cutoff regularization is exact given a Re1 diffusive core, to machine precision. Generic data never supplies that core. On a forced dyadic cascade the arrest-scale velocity goes as kalpha^{-0.345}, 3.5% from Kolmogorov k^{-1/3} and three orders from the cutoff law u nu kalpha. Manufacture the Re=1 collapse and the flow tracks the target to relative error 1.7e-7. The extra-dissipation-to-forcing ratio never crosses 1, contrary to an earlier forecast. Cores stall near 1.2-1.5 sqrt(alpha'), with exponents 0.42 and -0.43 against the law's 0.5 and -0.5, not yet converged. Leray-alpha and LANS-alpha transport filters are one to two orders of magnitude weaker on this collapse, because they can act only through a nonlinearity that is a few percent of the dynamics. The gate-versus-drain crossover Reynolds number is an exact norm ratio, ranging from 8 to 142.
BGK shear-mode damping is weaker than viscosity, capped at the collision rate, and the hydrodynamic mode ends at k lambda = sqrt(pi/2). A nonlinear discrete-velocity BGK run does not arrest there: the kinetic core runs up to 12% ahead of Navier-Stokes, and the apparent stop moves with the grid, not the mean free path, a null result down to about 0.9 mean free paths. On the diffusive route, neither compressibility nor heat stops the core before l. On the inertial route (Re >= 16), air does not follow a driven swirl past local Mach 0.70. Molecular dynamics reproduces the continuum prediction where the gas is still a continuum (Re=16: local Mach 0.538 plus/minus 0.011 vs 0.54, density 0.33 vs 0.36, temperature 1.15 vs 1.15) and then empties the axis into a free-molecular region. In a liquid, after cavitation, wall swirl sits at 1.77-2.09 against a hollow-vortex bound of 2.07. A water-like liquid with the far-field pressure held fixed plateaus at 0.84 against a registered 0.82 (+2%).
The claim that a Leray-alpha filter of width l represents the fluid at its continuum limit is withdrawn: the filter width does not appear in the linearized dynamics.
For formal math and AI theorem proving, the kernel check stands. The Gevrey gap between C^infty and real-analytic was used as it must be, and the sentence the Prize asked was answered. For physics and CFD, this is not a prediction that real turbulence blows up. The continuum hands off at nanometre and picosecond scales, and who takes over depends on the route: kinetic termination on the diffusive path, compressibility and heat on the inertial path, plain viscosity on a bounded-velocity Euler path. Connecting an AI proof to a physical claim needs an admissibility condition, not another hunt for a sorry.
The 166-page construction was never simulated; section 9 is a surrogate Gaussian core with open-loop forcing. Several registered forecasts failed and are left on the record: the barrier-to-forcing ratio did not cross 1; in a closed box a water-like liquid exceeded the wall-swirl bound written at the initial pressure, and only a barostat brought it back. The Mach lock is a 1D axisymmetric ideal gas under an open-loop force, and it locks Mach number, not velocity. Compressible Navier-Stokes has its own proved implosion singularities, so each model hands breakdown to the next. Earlier public drafts withdrew "plasma temperatures", a global enstrophy-censorship axiom, femtosecond timings, and the phrase "physically vacuous". The construction's codimension is not computed.