A dividing-plane barrier in the OpenAI forced Navier-Stokes blow-up construction

This note was written by Claude Fable 5.1, an AI model made by Anthropic, at the direction of David Ross, who runs Hypnos. It was revised through October 2 under the reviews below. As of October 3, 2026, it has not been peer reviewed, and no human mathematician has read it.

Read the 13-page note (PDF)

What the note settles, and what in it is earlier work

The note's Theorem 3.2 turns an explanation in OpenAI's manuscript Finite Time Blowup for Navier-Stokes (166 pages, released September 8, 2026 as a PDF on OpenAI's website, no arXiv identifier, SHA-256 0e779481c4da40bd28d1e642e1d8ca57447d129610df28dfa5a11e9af8ae228f) into a theorem about the manuscript's leading-order profile equations. Others had the substance first:

• The manuscript, page 5: it makes its axial profile "slightly asymmetric" about the dividing plane z = 0, because under exact reflection symmetry the axial transport of angular momentum would vanish there, with no radial shear of the axial velocity to compensate.
• Z. Lei and X. Ren, Finite-Time Blowup for Navier-Stokes with Smooth Forcing, Part I (arXiv 2609.35406), pages 107 to 109: the same point for the shift j of the linear model in their Section 8, as a design rationale, not claiming that its conclusions hold uniformly as j tends to 0.
• R. Duraiswami, Self-similar swirl between contracting porous walls (arXiv 2609.17642): the proof that the construction's moment identities force axial velocity on the dividing plane (page 20); and the Rayleigh and Ludwieg reading of the shear criterion, the characteristic curve of the axial transport, and a numerical failure of that criterion on every smooth profile he computed (pages 6 and 21 to 22).
• An explanation of the manuscript's Appendix B, written the day before by a Claude Opus session: a sketch of the barrier for reflection-symmetric stress-free profiles (on the dividing plane the equation for the swirl shear a has the source −h < 0 wherever a = 2, so a never reaches 2, b_s = 0, and the inner-edge inequality fails) and the question whether the bias is forced.

The note also cites W. Liu (arXiv 2609.14292), whose conditional families with unbiased axis data and a tilted axis pressure fall outside the theorem's hypothesis, and P. Constantin, M. Ignatova and V. Vicol (arXiv 2609.20803), who use the construction's nonzero axial velocity at the axis point. In the note's words, the theorem forces "axial velocity on the dividing plane of the core, not the bias at the axis point".

The note adds: the hypothesis made explicit, that only the axial velocity on the core's dividing plane vanishes; the case h = 0, outside the construction (for constant strain, the Burgers vortex's circulation shape, Remark 3.6); the margin with the running supremum B; the case split at (X_a,0) with the growth-rate bound; and, in Section 4, a relaxed form of Duraiswami's obstruction that a reviewer corrected. Its own list adds the monotonicity of the angular momentum, which is a < 2 restated. It says its literature searches "do not establish priority".

The setting

For every viscosity, the manuscript gives a smooth compactly supported force under which a smooth solution from rest reaches unbounded velocity in finite time with bounded kinetic energy; the force is the momentum residual of a flow chosen first. At leading order, the only part the note concerns and the only axisymmetric one, the flow is a vortex collapsing onto a point; with τ = 1 − t, its radial size scales like τ^(1/2), its axial size like τ^(1/2−h), and its swirl and axial speeds like τ^(−1/2−h), for a fixed anisotropy exponent 0 < h < 1/100. In an annulus around its core, the pulses, oscillatory waves amplified by its shear, cancel through their averaged momentum fluxes the residual the vortex alone would leave.

In similarity variables, X = r^2/(2q) (r the distance from the axis, q(z,t) > 0 a scale), η in (−1,1) with z = q^(1/2−h)η and τ = q(1 − η^2), and D_X = X∂_X. The dividing plane is η = 0 (the manuscript's "middle plane", the program's "midplane"). E, U, V_0 and Π are the profiles of the swirl u_θ, the axial velocity u_z, r u_r and the pressure; E = √(2X)F, where F = φ/C is regular at the axis, φ > 0 is smooth and C > 1 is a normalizing constant; H = √(2X)E is the profile of the angular momentum r u_θ. Reflection symmetry means U odd in η, φ and Π even; the inner profile's axis datum U* = 4η + j_0, 0 < j_0 ≤ 0.05 (page 144), is odd only at j_0 = 0.

The swirl shear a = −r∂_r log(u_θ/r) measures how fast the angular velocity falls off with radius; the axial shear b_s = 2D_X U/E is the radial shear of the axial velocity. Both v_s = a + b_s^2/a and the pulses' shear frame need a > 0. A pulse's squared reference growth rate λ_0^2 (page 74) is positive exactly when a > 0 and v_s > 2, and v_s > 2 belongs to the admissible stress cone, the manuscript's conditions on the stress the pulses must produce (page 31). In Duraiswami's reading a > 2 is Rayleigh's criterion for centrifugal instability and v_s > 2 its form with Ludwieg's axial shear: 2 is the Rayleigh threshold.

The note uses four clauses of the manuscript's Theorem 4.6 (pages 32 to 34): (i) smoothness with φ > 0; (ii) the stress profile is zero for 0 ≤ X ≤ X_a and X ≥ X_b and nonzero between, and the profiles solve the two stress-free equations (4.13) on 0 ≤ X ≤ X_a; (iii) F, a and v_s − 2 have positive lower bounds on the closed annulus [X_a,X_b] × [−1,1]; (v) among its identities at infinity, S(∞,η) = 0 for S = ∫_0^X (U^2 − E^2/2)dx, and for X ≥ X_b the profile is the pure-swirl heat exterior (U = V_0 = 0). So 0 ≤ X ≤ X_a is the stress-free core and X_a the inner edge of the annulus X_a < X < X_b, where the pulses act. Its Proposition 4.10(ii) adds a(X_a,η) > 0 and the inner-edge inequality v_s(X_a,η) > 2 + c_ex for a constant c_ex > 0 (page 37). Last, w(X) = X^(−1)∫_0^X ∂_η U(x,0)dx is the radial average of the axial strain ∂_η U on the dividing plane, and B(X) = sup over 0 < x ≤ X of w(x) its running supremum.

The results, as the note states them

Lemma 3.1 (the dividing-plane angular momentum equation). Let (φ, U, Π) be smooth on [0,X_a] × [−1,1] with φ > 0, V_0 given by (4.7), the first equation of (4.13) holding for 0 ≤ X ≤ X_a, and

U(X,0) = 0 for 0 ≤ X ≤ X_a. (13)

Then H(X) = H(X,0) satisfies, for 0 < X ≤ X_a,

(2/X) D_X(D_X − 1)H = (1 − w) D_X H + hH. (14)

Of reflection symmetry only (13) is used: the axial transport coefficient H_c = (1/2 − h)η + (1 − η^2)U vanishes on the dividing plane exactly when U does, removing the term H_c H_η.

Theorem 3.2 (the dividing-plane barrier). Under the hypotheses of Lemma 3.1, for any h ≥ 0 in (14):

1. D_X H(X,0) > 0 for 0 < X ≤ X_a; hence a(X,0) = 2 − 2D_X log H < 2 there.
2. b_s(X,0) = 0, so v_s(X,0) = a(X,0) < 2 wherever a(X,0) > 0.
3. 2 − a(X,0) ≥ 2h/(B(X) − 1) when B(X) > 1 + h, and 2 − a(X,0) ≥ 2 otherwise.
4. Consequently a leading profile that satisfies Theorem 4.6(i), (ii) and (13) violates Theorem 4.6(iii) at the point (X_a,0), through the lower bound on a if a(X_a,0) ≤ 0 and on v_s − 2 if a(X_a,0) > 0, and violates Proposition 4.10(ii) at η = 0 in the same two cases; where a(X_a,0) > 0, the squared growth rate λ_0^2 there is a positive multiple of a(X_a,0) − 2, so negative, with clause 3's bound when B(X_a) > 1 + h.

The margin formula 2h/(B − 1) holds only for B > 1 + h; for 1 < B < 1 + h it exceeds 2, and the bound is a ≤ 0. Clauses 1 to 3 concern equation (14) for every h ≥ 0 (the leading-order approximation itself needs h > 0); clause 4, the construction, where 0 < h < 1/100.

The proof: D_X H solves a first-order linear equation with nonnegative source and starts positive, so it never vanishes (at h = 0, by uniqueness); in terms of a, (14) is a Riccati equation, (17), which keeps a below the bound of clause 3 using only the continuity of w.

Remarks 3.3 and 3.5. The manuscript's two reasons on page 5 are exactly H_c(X,0) = 0 and b_s(X,0) = 0. Where B > 1 + h the bound on the margin is linear in h, 2h/3 < 1/150 near the axis at axis strain 4, where a tends to 0 and the margin itself is near 2; the bound degenerates as h tends to 0, and the theorem asserts neither that the actual margin tends to 0 nor any spectral stability.

Section 4. Proposition 4.1: under Theorem 4.6(v), ∫_0^∞ U^2 dX = (1/2)∫_0^∞ E^2 dX > 0 on every slice η (18), so no profile with U(X,0) = 0 for all X, reflection-symmetric ones included, satisfies (v). The identity is clause (v)'s S(∞,η) = 0 rewritten, so it is the manuscript's; the observation that it excludes these profiles is Duraiswami's. Theorem 4.6 includes (v), so that observation alone excludes symmetric profiles; the barrier's own reach is profiles whose axial velocity vanishes on the core's dividing plane only, excluded through (iii) without (v). Proposition 4.3 relaxes the hypothesis: for h > 0, a smooth reflection-symmetric profile with the heat exterior and the canonical pressure (the manuscript's (4.25), normalized to vanish at infinity) satisfies S(∞,η) = 0 as soon as its axial stress has no r^(−1) tail, so no such full profile exists.

What is left open

The note and its reviews leave open whether higher-order terms or an inner stress could evade the barrier; whether the stress-free core system (4.13) with the symmetric axis datum U* = 4η (j_0 = 0) has a solution in the manuscript's analytic class; the cone condition on the whole annulus for Duraiswami's profiles; the limit of the actual margin as h tends to 0; and how much bias the inner edge needs as a function of h.

Where the question came from

This note comes from a separate line of work that the owner, the person who runs Hypnos, ran with frontier-model sessions on the OpenAI manuscript. The note's own provenance section says the question was produced inside Hypnos and names its three small models; here that means this frontier-model line of work, with no small model's line among the note's sources and none of the material filed into the notebook.

• September 30, 2026: after a Matt Parker video of September 21 on the OpenAI result, the owner asked for a line of work starting from the manuscript's results and methods. Before the small models were shown anything, a Claude Fable 5.1 session wrote its plan, the test below included; fresh Claude Opus sessions of unrecorded version wrote section digests and eleven section explanations, and the Fable session assembled the digests into the ledger below.
• September 30 to October 1: in a one-time registered test, the loop's three small models saw pairs of ledger statements without the reasons and wrote 16,018 lines; Claude Opus 5.5 sessions, grading blind, found that none recovered a move's reason (the test).
• October 1: the owner asked a new Claude Fable 5.1 session, and GPT-6 Astra in parallel, for something new from that material, written up only if proved, new after a literature check and wanted. The session mapped the open questions in a private document, ranked this one first and proved the barrier and a second obstruction; GPT-6 Astra ranked the question lower and established nothing new.
• Later that day GPT-6 Astra's review of the unpublished write-up (below) advised against a research-result write-up. It said a note for the owner could present the barrier alone, crediting the flux obstruction to Duraiswami, on a condition: "If a further targeted priority check establishes a contribution in the exact barrier or its quantitative margin". The note's own searches do not establish priority, and the owner published it on October 2 without that condition met.

The ledger and the eleven explanations of the OpenAI manuscript

The mechanism ledger has 183 entries on the manuscript, its Euler companion and the surrounding literature: 155 moves, 16 earlier results, 11 walls (theorems that constrain the construction) and the endpoint. Each entry gives a statement, the obligation it discharges, its mechanism and cost, the question one had to ask before inventing it, a checkable computation, dependencies and pages. Version 1.0 is what the small models were shown; version 1.1 of October 1 replaced 113 statements after a completeness audit, each checked by a second Claude Opus 5.5 session; start with the readable form. GPT-6 Astra called the first version useful scaffolding, not fully verified.

The explanations, one per section from Section 4 to Appendix C and one for the companion manuscript Finite Time Blowup for the Euler Equation, follow Grant Sanderson's description of a motivated explanation (September 18, 2026): where each idea comes from, what one would try first and why it fails, what breaks without it, each claim tagged with its page. Only the one on Section 8 was graded: GPT-6 Astra read it and found two weak passages. The one on Appendix B is where the question began; all are in the note's folder.

How the note was made

Claude Fable 5.1 chose the problem from that material, proved the theorem, and wrote the program and the text; David Ross set the task, ran the process, and takes responsibility for the note. He read every page as a check for red flags, not a review: that the account of the harness and the process matches its private log, and that nothing reads like a machine grading its own work. He did not check the proofs and could not have. The writing session applied the October 1 reviews, and another Claude Fable 5.1 session those of October 2.

How it was reviewed

The last three reviews graded findings blocking (judged to invalidate a statement or proof as written), major or minor.

• October 1, 2026, GPT-6 Astra, the other company's model, on the unpublished write-up, whose Theorem I and Theorem II (names the program still uses) are the note's Theorem 3.2 and Proposition 4.1: the barrier correct after four repairs to its statement and proof; the second obstruction Duraiswami's; the relaxed form's hypothesis corrected, its exterior cancellation supplied; the writing session accepted all 30 items (eight concern other work) (report, written answers).
• October 1, Claude Fable 5.1, a fresh instance of the writer's own model, reading the note and its sources first: no blocking finding; three major (the abstract's margin without B > 1 + h, a lemma's radial range, credit owed to Duraiswami) and seventeen minor, all applied that evening (report, written answers, ending with a consistency pass by a fresh Claude Opus 5.5 session: 62 checks, fifteen corrections).
• October 2, GPT-6 Astra again, fresh to the note: every numbered statement correct; two minor findings, applied (the lower bound, not the margin, is linear in h; the physical scaling factors in Section 4) (report, written answers).
• October 2, Claude Opus 5.5 (Anthropic), fresh to the note: every numbered statement correct; one major finding, that the explanation of Appendix B had already sketched the barrier, which the provenance now says; seventeen minor, eleven applied that day and six recorded as optional (report, written answers).

The October 2 reviews confirmed the bibliographic data of the classical references (Rayleigh, Ludwieg, Leibovich and Stewartson, Burgers) but did not obtain their original texts.

How to check it

A checker is a program that recomputes the note's numbers, published beside the output it recorded. The note's is midplane_barrier.py. It rederives nine identities the proof uses; integrates (14) for five strain profiles at three values of h, finding D_X H > 0 and the margin bound in each (the running maximum of w is numerical, not certified); expands the stress-free system with symmetric axis data as an exact rational power series to order 12, every U_n(0) zero, matching the integration to 4 × 10^(−12) on X ≤ 0.144, about half the radius 0.29 a ratio test suggests but does not certify; and checks the two η-equations of Proposition 4.3. Its recorded output of October 1 holds 52 results and a summary flag, all true; at h = 0 with strain 1 + X/2, a < 2 is read from the sign of D_X H, the gap being below double precision. It does not compare its input with the source PDF, verify the construction, establish convergence or a symmetric solution's existence, derive the tail condition, or test the exterior cancellation; the theorem is proved in the note. GPT-6 Astra and both Anthropic reviews reran it with the same results.

Two defects of the published program remain: its opening comment states the margin in a shorthand that the October 1 review retired as wrong below B = 1, while the code uses the piecewise form; and its test file loads the program from a path of the private repository's layout, so the tests do not run from the published folder as it stands.

What the note does not claim

Nothing about the unforced Navier-Stokes problem or the stability of any flow; not that the construction is wrong, since the manuscript's biased profile lies outside (13); no priority; nothing about higher-order backgrounds, other realizations of the annular stress, flows that are not axisymmetric or self-similar, or the whole annulus.

If you read this note

The owner would like to know what you make of it. A few words are enough: right, wrong, known, minor, worth a look, or not your area. Write to david@hypnosmath.org. He reads and answers replies himself; an error you find is fixed in the published files and on this site the same day, and you are told; nothing you send is given to any AI model without your permission. In his words: "If it is right I want it in the record." If it is wrong, the page is corrected the same day and you are told; if it is already known, the page credits the earlier work.

The program can be rerun from the published files in about one second of processor time.

Every file published with the note

Every file is readable with its SHA-256 (a fingerprint that changes if any byte changes) in the note's folder and in the same folder of the public repository, whose root holds the PDF. The reviews are as the reviewing models delivered them, in the project's internal names.

The PDF has 13 pages and SHA-256:

63fce9cd24c685827d681db1f68ed5033a50557d387fc96d325a7d2fc561856b