The OpenAI Navier-Stokes and Euler Blow-Up Proofs: A Physical Reading, Not a Physical Refutation

Socrate AI Lab / MechanicaFluidorum Program · Lead: Xavier Callens GitHub: xaviercallens/OpenAI-NSE-Epistemic-Audit (release v5.5.0) Zenodo: concept DOI 10.5281/zenodo.22696717 (always resolves to the latest version) · v5.6.0: 10.5281/zenodo.22838708

What this is

In September 2026, an OpenAI multi-agent system produced Lean 4-verified proofs of finite-time blow-up for the forced 3D Navier-Stokes equations (Millennium Prize Alternatives C and D) and for the unforced Euler equations: zero sorrys, zero custom axioms, no weakened norms. This dataset does not dispute that proof. The Millennium Prize problems ask precise questions about a specific continuum model; they were never claims about how real fluids behave. This project is a physical reading of the specific solution OpenAI constructed: how far it travels from the regime the incompressible model is normally trusted to describe.

Headline result

Because the collapsing core keeps a radial Reynolds number of order one, its scales are diffusive (ℓ_r ≈ √(νt), u ≈ √(ν/t)), essentially independent of the initial vortex size. Compressibility, rarefaction and viscous heating all become order-one effects at a single length ℓ* = ν/c_s (0.7 nm in water, 45 nm in air), a few picoseconds (water) or nanoseconds (air) before the mathematical singularity -- expressible as one local vorticity bound |ω| ≲ c_s²/ν that the Beale-Kato-Majda theorem turns into a genuine admissibility criterion. For liquids, cavitation is reached three decades earlier still.

Results added in v5.2.0 to v5.8.0

• 3D solver validated. Pseudo-spectral Navier-Stokes solver reproduces the Taylor-Green Re = 1600 benchmark: dissipation peak at t = 9.14 against the published t = 9.0.
• Cutoff law tested. Exact given its premise, but the premise (a Re ~ 1 diffusive core) is not produced by generic data: in a cascade the arrest-scale velocity scales as k^-0.345 (Kolmogorov -1/3), where the law predicts k^+1.
• Forced collapsing core. A manufactured Re = 1 collapse is tracked to 1.7e-7; the barrier's engagement collapses onto the single variable alpha'/(nu tau) (per-run prefactor 0.188 +/- 0.010).
• 96^3 sweep (v5.4). The forecast that barrier dissipation would overtake the forcing (B/F = 1) failed: 0 of 6 runs cross. The core instead stalls at l ~ 1.2-1.5 sqrt(alpha') (collapse rate 0.04-0.16 against 0.500 without barrier), with window-end exponents l +0.42, u -0.43 against the law's +0.5, -0.5. Not yet converged.
• Gate vs drain. On that collapse, Leray-alpha / LANS-alpha lag it by 0.2-0.4%, a hyperviscous barrier by 5-45%. The crossover Reynolds number Re_x = ||(L_barrier - L_nu) U|| / ||N_alpha(U) - N(U)|| ranges 8-142. A ranking of two models.
• Kinetic anchor. l*/lambda = cbar/(2 c_s) = 0.67 for air: the validity scale is the mean free path, with a derived constant.
• Kinetic lock, linear (v5.3.0). The exact BGK shear-mode spectrum damps small scales LESS than viscosity (Gamma = nu k^2 [1 - (k lambda)^2 + ...]), never exceeds the collision rate 1/tau, and the hydrodynamic mode ceases to exist at k lambda = sqrt(pi/2) = 1.2533.
• Kinetic lock, nonlinear (v5.4) -- a null result. A validated 2D discrete-velocity BGK solver in Rust, driven by the same manufactured collapse, does NOT arrest it near k lambda = sqrt(pi/2): the kinetic core runs up to 12% ahead of Navier-Stokes, and its apparent stopping point moves with the grid (0.98 -> 0.52 lambda at Re 1 as dx goes 0.67 -> 0.17 lambda) while a Navier-Stokes control on the same grid tracks its target to 4e-4. Robust down to ~0.9 lambda; isothermal, 2D.
• Lean 4. 85 declarations in ten files on the three standard axioms: scaling chain, Leray-alpha filter bounds, discrete and nonlinear BGK H-theorems, the kinetic eigenvalue cap -1/tau <= Re mu <= 0, the gate/drain energy identity, and -- with no remaining hypothesis since v5.4.0 -- the statement on OpenAI's own ProblemStatement objects that any object with their CandidateProperties exceeds every velocity-gradient bound arbitrarily close to t = 1.
• Regime map (v5.5.0). Kn = Ma/Re sorts blow-up scenarios by the physics they meet first: OpenAI's construction (Re ~ 1) meets everything at l*; Tao's averaged-equation blow-up (Re -> infinity) meets compressibility first, inside the continuum; a human-written forced Euler blow-up with bounded velocity meets viscosity first. Proved in lean4/BlowupRegimeMap.lean.
• Compressible / thermal forced core (v5.5.0). On OpenAI's route neither compressibility nor heat stops the driven core before l* (lag 14-25%; an arrest prediction of ours failed). On the inertial route (Re >= 16) air locks at local Mach 0.70 however far the target is driven -- a lock on Mach number, not on velocity or size; 1D ideal gas, open-loop force; the compressible equations have their own proved implosion singularities. code/compressible_core.py, data/compressible_core_*.json.
• Leray-alpha anchor claim withdrawn (v5.5.0). alpha does not appear in the linearized dynamics of any alpha-model (lean4/LerayAlphaLinearization.lean), so it cannot be calibrated to l*.
• Closure-free test of the Mach lock (v5.6.0-5.6.1). Molecular dynamics of a Lennard-Jones gas driven by the same force, viscosity measured in situ, against a continuum prediction registered before the MD data existed: at Re = 16 and target Mach 1, local Mach 0.56-0.62 vs 0.54, core density 0.34 vs 0.36, core temperature 1.15 vs 1.15 (v5.6.1: three runs, 0.538 +- 0.011 vs 0.54 with a consistent real-gas sound speed; Re 32: 0.62-0.71 in one run, 0.64-0.73 in two). Beyond target Mach ~1.5 the axis becomes free-molecular. In a liquid the core cavitates and the swirl at the cavity wall levels off at 1.77-2.09 (two runs), at or below the hollow-vortex bound 2.07 within one standard error.
• Three-dimensional boxes, a water-like liquid and a failed prediction (v5.7.0). 3D gas: local Mach <= 0.87 while the target passes 3, no axial instability above an undriven vortex. Re = 4 (3 runs): 0.29/0.40/0.50 vs continuum 0.27/0.39/0.48. Water-like liquid: the wall swirl exceeded the bound registered at the initial pressure (0.94-1.14 vs 0.82: that prediction failed); per-bin pressure (md_run --stress on) shows the closed box's ambient pressure rose 0.32 -> 0.77, and at that pressure the swirl is 0.45-0.75 of the bound (post hoc, one run). v5.8.0: with a barostat holding the far-field pressure fixed (--baro auto, two runs) the wall swirl plateaus at 0.84 vs the registered 0.82 (+2%, within one standard error): the prediction held at fixed ambient pressure. Raw runs in data/md_core_runs/, code in code/md_core_rs/. data/md_core_*.json, figures/md_core.png.
• Quantum-fluid counterpart (v5.6.0). A quantized vortex is a Reynolds-number-one core by theorem; quantum pressure, heat and cavitation all let a fluid carry a velocity singularity by emptying the core. research/QUANTUM_FLUID_MICRO_MACRO_LINK.md, lean4/QuantumVortexLink.lean.
• Reproducibility benchmark (v5.4.1). BENCHMARKS.md; code/run_benchmarks.sh re-runs tests, Lean files, solver gates and fast simulations and checks every regenerated number against the data.

What changed from earlier releases (important)

An earlier version of this dataset (still visible in this repo's commit history) asserted a different framing: a global "Thermodynamic Censorship" enstrophy axiom, "plasma temperatures", and a Mach-limit timeline that misread the model's own dimensionless time parameter as seconds directly (off by twelve orders of magnitude -- femtoseconds instead of picoseconds). Those claims are withdrawn. See CHANGELOG.md and Appendix A of the paper for the itemized list, and PEER_REVIEW_2026-09-15.md for an open peer review and the authors' point-by-point response. superseded/ in this repo holds the retracted paper and Lean file, kept for the historical record with inline withdrawal notices -- do not cite them for their original claims.

Contents

• paper/: current flagship paper (PDF + LaTeX source)
• workstreams/: six community-research-direction notes (admissibility, turbulence-modeling closure, thermal response, cavitation, divergence-free vs. incompressible, genericity/codimension)
• scripts/: the two verified analytical scripts referenced in the paper's physical-scale tables
• outputs/: their console output logs (regenerated 2026-09-15; times are explicitly labelled dimensionless τ vs. physical seconds t = Tτ)
• superseded/: the retracted paper and Lean file, with withdrawal notices, for the historical record
• lean4/: the ten verified Lean 4 files (85 declarations) and their README (standard axioms only)
• research/: programme notes and experiment write-ups
• code/: every simulation used in the paper -- code/*.py (3D pseudo-spectral solver, forced-core test beds, 1D compressible Navier-Stokes-Fourier core, Gross-Pitaevskii solvers, BGK spectrum, analysis scripts), code/tests/ (the pytest suite), and two Rust crates: code/kinetic_lock_rs/ (discrete-velocity BGK solver) and code/md_core_rs/ (molecular dynamics of the forced core); code/benchmark/ re-runs everything
• figures/, data/: every figure and JSON result of the experiments (v5.2.0 onwards); data/md_core_runs/ holds the raw molecular-dynamics runs (validation gates, viscosity measurements and every forced run)
• CHANGELOG.md, PEER_REVIEW_2026-09-15.md, PROJECT_README.md: project documentation

Citation

Cite the GitHub release or the Zenodo record above, not this dataset card directly. Please do not cite superseded/Thermodynamic_Censorship_Navier_Stokes.pdf for its original conclusions.