The Proof Is in the Vault: OpenAI, Anthropic, and the Negotiated Release of Machine Science

On Monday, October 26, OpenAI announced that its internal models had produced three discoveries: a proof of the local smoothing conjecture — a central problem of harmonic analysis — in every spatial dimension for the first time; a Lean-checked refutation of the Exponential Time Hypothesis, handed to a verification panel that includes the two mathematicians who proposed it in 1999; and a rigorous proof that certain quantum field theories — Yang-Mills and QCD among them — exist in four dimensions. The math was dumped on GitHub raw. The computer science got referees. The physics — in natural-language form since mid-September, formalized in Lean by October 16 — was not released at all, deferred to external “shepherds” who will decide how the world gets to see it. The Advisory Group on Mathematics and AI immediately demanded its release, noting OpenAI’s own statements say formalization is complete. Hours after OpenAI’s announcement, Anthropic published its own claimed ETH refutation. Run it through four frames.

1. Assumption Audit: The Field Ran on an Unchecked Premise

The Exponential Time Hypothesis holds that the hard logic problems behind k-SAT necessarily scale exponentially, no matter how clever the algorithm. It is the load-bearing wall under hundreds of papers of the form “assuming ETH, no algorithm can do better than…” — conditionals that would survive individually if it fell, but would redraw the map of what computers can be made to do. The audit’s irony is structural: the field held the premise for 27 years on confidence, not proof, rating its truth high because nothing better existed — exactly the failure mode this recipe exists to catch.

Now audit the release protocol stacked on top. AGMAI’s demand is itself an audit: OpenAI says formalization is done; done formalization means the result is checkable; therefore whatever remains holding the physics result back is not scientific. OpenAI’s counter-assumption is definitional: release is a discretionary act rather than an obligation, and readiness is the lab’s to define. That is the keystone — that the tempo at which humanity learns what its machines have learned may be set by a private company. The cheapest test of the keystone costs nothing: publish the Lean files. The vault’s contents cost millions to produce and nothing to open.

2. First-Principles Decomposition: A Discovery Is a Supply Chain

Strip “discovery” to components: construction (a candidate proof), verification (a checker confirms the derivation), understanding (someone can explain why it works), credit (who owns it), use (what it enables). The traditional pipeline bundled all five in one skull or one lab — which is why priority disputes were so bitter: first meant everything. Machine science splits the bundle. OpenAI’s model now supplies construction and machine-verification at industrial scale — three fields in one announcement, weeks after the 722-manuscript dump of October 6 and the quiet withdrawal of three of those manuscripts for a sign error. Understanding lags: Ryan Williams, one of three ETH referees, has not certified the result and told TechCrunch the proof contains “two or three alien-looking parts” — verified, not yet comprehended. Credit is contested: Alpöge and Sussman were, by their own account, weeks from an independent QCD proof and updated their public repository the night before OpenAI’s announcement. Use remains theoretical — Williams again: if true, P=NP deserves “a little more weight”. Rebuilt from fundamentals, a discovery is not an event but a supply chain, and every link except one has been automated or outsourced. Timing is the lab’s monopoly.

3. Nietzsche Ladder: Camel, Lion, Child

Camel. Carry the inherited account first. OpenAI’s stated protocol is, on its face, generous: “we do not seek academic credit”; mathematicians got raw output to organize their own explanations; computer scientists got referees; physicists get a negotiated release. The Camel also bears the community’s real grief, which Altman narrated as a five-day arc from “this field is over” to acceptance — as though spectating a funeral were the same as attending one.

Lion (responding to Camel). You have carried the tablets well — now ask who wrote them. A lab that seeks no credit is collecting altitude: “AI makes discoveries” headlines land the same week Oracle folds OpenAI models into the enterprise stack, days after its rival’s S-1 primed the market — and both labs also claimed cryptography breakthroughs the same day. The shepherd system converts the community from judge to supplicant: computer scientists now petition the lab for raw data and request that OpenAI fund their own conference. And priority — the one currency science still mints — was repriced by announcement scheduling when Alpöge and Sussman’s night-before repository update counted for nothing. No: the gift has a gate, and the gate has a product roadmap.

Child (responding to Lion). Your No clears space; build in it. Verification is now nearly free — Lean checks what skulls cannot — so the commons no longer needs the lab’s permission to check, only to possess. A child’s protocol writes itself: formalization files land in public repositories at the moment of internal completion; release becomes a timestamp, not a negotiation; and credit flows to whoever explains the alien parts first, because explanation is the only human work left standing. Science spent centuries separating the discovery from the discoverer’s status. The machines offer the same gift — if the intake valve is public.

4. Question Forge: Stop Asking If the Proof Is Correct

The question consuming every timeline — is the ETH refutation correct? — is operational. Lean answers it on its own schedule, and asking it lets the announcement set the agenda; it is a shield question standing guard in front of harder ones. The diagnosis: it displaces the subject. The proof’s correctness is the artifact’s problem; the community’s wobble is about ownership. The forged question: when a private lab can hold a completed proof of an existence theorem about nature for six weeks — and the scientific community’s only move is to petition for its release — what does the public still own? What changed: the subject shifts from artifact to ownership, the question implicates lab and community alike, and answering it honestly costs something — it requires admitting the commons is now a customer. Carry it. Do not expect it to resolve.

What the Frames Converge On

Not one frame turned on whether ETH is actually false. Fomin called the claim “a complete surprise, a shock”; Goldberg’s “enormous breakthrough” is conditional on confirmation; the referees themselves have not certified. The frames agree on the fault line underneath: an assumption held for 27 years proved auditable in a day — and the auditor decides what the public sees, when, and in what order. As Hannah Fry put it, this should have been a glorious moment for mathematicians; instead, a massive collective wobble. Watch two testables: whether AGMAI’s petition springs the physics files from the vault, and whether the computer scientists’ requested conference gets funded by the lab they petitioned. The discovery may prove historic. But the proof is in the vault — and the vault has a release calendar.