8Braid pins an exact margin inside OpenAI's fluid proof

The number to hold onto is 7/8. Working inside one algebraic step of OpenAI's September 8 Navier-Stokes release, the independent team at 8braid proved in Lean that if the target stays inside its allowed cone with margin m, where 0 < m < 1, an error of at most m/4 in each of the four matrix entries keeps the correction weights positive and the determinant at 7/8 or above.

At a glance

• 8Braid re-ran the submitted formal checks first: both the Nanoda kernel and Lean's default kernel accepted the solution, and the recorded axiom reports named only the three standard Lean axioms.
• The new bound covers Proposition 7.5 on printed pages 82 to 83 and holds across the stated real-parameter domain rather than at sampled points, with separate Python checks rebuilding the rational algebra.
• Applying the allowance to the full construction still needs the source's analytic error envelope and profile bounds, and the database workload never ingested OpenAI's complete construction at all.

If you missed the original release: OpenAI published an AI-generated proof that a fluid starting at rest, driven by a smooth force, can break down in finite time, with a Lean formalization attached. According to alphaXiv, OpenAI said an internal model group got there in 88 hours using around 10,000 coordinating agents. The day before, Science News reports, Tristan Buckmaster of NYU and Levent Alpöge of Anthropic had announced an advance on the forced Euler equations with AI help.

Nanoda and Lean's default kernel both accepted the submitted proofs

8Braid started where anyone would, with the release itself. Its Comparator workflow matched the submitted C and D results to their expected formal statements, and two kernel checkers, Nanoda and Lean's own default kernel, accepted the solution. The recorded axiom reports named the three standard Lean axioms and nothing else: propext, Classical.choice and Quot.sound.

The team also kept the dull metadata that makes a replay possible: upstream revision f9e8bc5b38b6, dependency pins, checker identities, the commands run and the outcomes they produced. The successful Linux run kept the original challenge configuration and the genuine sandbox. 8Braid says plainly that this establishes one thing, that the formal proofs were accepted against the encoded statements, and that formal reproduction is neither specialist peer review nor a decision by the Clay Mathematics Institute.

Worth keeping in view: the full OpenAI manuscript runs more than 150 pages, and as of the Science News report no one had completely verified it, certainly not on the human side, in the words of mathematical physicist Gregory Eyink of Johns Hopkins.

Proposition 7.5 needed room for error, and now it has m/4

Then 8Braid went inside a single construction step. The construction as a whole, per Wolfram Community's description, is a vortex whose core shrinks while its velocity grows, with the driving force staying smooth. Proposition 7.5, equations 7.24 to 7.28, uses a positive covariance decomposition: two contributions mixed with weights that must stay positive. The open question is how far the inputs can move before that positivity breaks.

The answer: with the normalized matrix written as [[1+a, 1+b], [-1+c, 1+d]] and the target at (1, s) with |s| ≤ 1−m, errors of at most m/4 in each of the four entries preserve the positive correction weights and hold the determinant at 7/8 or above. The fractions are exact, and the result covers the stated real-parameter domain rather than a collection of sampled points.

Python checks rebuild the rational algebra; a standalone Lean proof checks the general implication. 8Braid claims no optimality for the allowance and does not instantiate the source profiles, prove their error envelope or settle smooth extension at a zero target. It is careful to say the original step was not defective for leaving a small constant abstract; what it adds is an explicit conditional bound with assumptions attached.

The withdrawal test ran on a fixture of 88 facts and 26 rules

The conditional lemma then went through a native 8DB research workload, which tracks whether a result is available alongside the obligations needed to apply it. 8Braid withdrew the supporting admission and read the saved state again. The covariance lemma stopped being recorded as available for its declared scope, the application to the actual source construction kept its separate unadmitted obligations, and unrelated historical results kept their support.

Withdrawal changes the record of support, not the truth of the mathematical implication. The test passed through the native consumer and a separate reader, including a replay from a fresh local directory, and same-host relocation was checked. The fixture holds 88 facts and 26 rules; seven adapter tests and 24 refusal controls passed; the six source and PDE obligations stay unadmitted, and an earlier qualified 1/73 result is preserved.

A checker sits between the agent's proposal and the record

The design being tested puts a checker between what an agent proposes and what enters the research record. A changed polynomial coefficient fails an exact identity check. An artifact carried over from the wrong context fails an applicability check. An inference with missing support stays unadmitted, and the agent gets a concrete reason to revise its next step rather than a vague failure.

Think of it as a receipt drawer rather than a filing cabinet: each claim sits with the slip saying who checked it, against which statement and under which assumptions, so pulling one slip empties one claim and leaves the rest alone. Each tool has one job. Lean checks formal proofs, exact arithmetic supplies checkable algebraic certificates, and the 8DB workload connects evidence to the claims that depend on it.

The limits are stated, and they bite. The workload used a fixed Taylor-Green research context, never ingested OpenAI's full construction and never independently certified the PDE, and 8Braid has measured neither a reduction in an end-to-end agent's error rate nor a production latency guarantee. Oddly, the same work that pins an exact 7/8 floor cannot yet say which source parameters satisfy the condition, because the analytic error envelope and profile bounds are missing.

Which source columns come next

8Braid's stated next mathematical task is supplying the analytic bounds needed to apply the covariance allowance to selected source columns; the database task is moving the result into a context-independent theorem and evidence workflow. The named experiment: enclose one set of source parameters with VTTE's interval-backed tools, check the covariance margin, then check agreement where two certified parameter regions overlap.

No date is given for either. Per Wikipedia, the Clay Mathematics Institute still lists the Navier-Stokes problem as active as of September 2026.