Astra Solved 10 Open Problems. The Lean Kernel Is the New Authority.

On August 1, OpenAI pushed a repository called ten-proofs to GitHub. It contains Lean 4 certificates for ten previously-unsolved mathematical problems: the first explicit construction of a non-sofic group, a disproof of Connes’s rigidity conjecture, improvements to sphere-packing bounds, three resolutions from the Erdős problem catalogue, and four more. The repository metadata includes one figure that matters more than any of the problem names: sorry count zero. Every formal step in every proof is spelled out. Nothing has been left as an exercise for a later mathematician.

This is not the same as saying AI did math. AI has been doing math — plausibly, impressively, inconsistently — for long enough that the community has developed reflexes for evaluating claims. The reflexes exist because AI math claims have failed before. In October 2025, OpenAI vice president Kevin Weil announced that the same model family had solved ten Erdős problems. Thomas Bloom, who maintains the canonical Erdős problem registry at aiweekly.co, called that announcement “a dramatic misrepresentation.” The model had found references to existing solutions in the literature, not produced new mathematics. The embarrassment was real and specific, and it established what the standard should be: proofs with receipts. Astra produced the receipts.

What Zero Sorries Means

Lean is a proof assistant, not a language model. It is a formal system in which mathematical statements are encoded as types and proofs are programs that inhabit those types. When a mathematician writes a proof in Lean 4, every inference step must be fully specified — the system does not accept “and clearly this follows” because it cannot evaluate the word “clearly.” The Lean kernel, a small C++ core of roughly 20,000 lines, type-checks the resulting program. A proof either compiles or it does not. The sorry keyword is how Lean handles incomplete work: it lets you mark a gap in the argument, advance past it, and come back later. A proof with sorries is a proof with holes. As Pondero’s coverage notes, “A Lean proof that type-checks is valid by construction. No trust in OpenAI’s evaluation claims, benchmarks, or interpretation is required.”

This is a structural change in the relationship between AI and mathematical truth. Every AI math result before it — including AlphaProof’s IMO solutions, including GPT-5.6’s CDC conjecture work — required either a human mathematician to read and validate the argument or, in the case of Lean-assisted proofs, a human to do the formalization. Astra’s certificates were formalized with human assistance in manuscript preparation, but the mathematical reasoning — the identification of what to prove and how — originated from the model. SiliconAngle’s coverage notes that a 249-page manuscript collection was published alongside the repository, containing “model-written reasoning walkthroughs and Lean 4 certificates.” The walkthroughs are the AI’s mathematical thinking made legible; the certificates are the proof that the thinking was correct.

What the Lean kernel returns is binary: valid or not. Anyone with Lean 4.32.0 installed and the repository cloned can run lake build All and watch all ten proofs type-check. No trust in OpenAI’s evaluation team, no trust in the model, no expert endorsement required. Thomas Bloom — the same mathematician who called the October 2025 claims a “dramatic misrepresentation” — called these results “big news,” and rated them more significant than OpenAI’s earlier unit distance conjecture disproof. That reversal in Bloom’s position is worth registering. The evidence is different this time.

The Ten Results

The headline problem is the non-sofic group construction. Mikhail Gromov introduced the concept of soficity in 1999, describing groups that can be well approximated by finite symmetric groups — that is, groups where every finite subset of elements can be faithfully represented in a finite permutation group. The question of whether any non-sofic groups exist was left open. Gromov’s intuition was that perhaps all groups were sofic; if so, a large body of theorems that hold conditionally on soficity would become unconditional. Astra constructed an explicit example of a non-sofic group, ending 27 years of conjecture. The certificate is in NonSoficGroup.lean.

The second result that will command serious attention is the disproof of Connes’s rigidity conjecture. The conjecture, from the theory of operator algebras, concerned whether a group’s von Neumann algebra uniquely determines the group. Connes believed it might. Astra constructed two non-isomorphic groups that produce the same von Neumann algebra — a counterexample formalized in ConnesRigidity.lean. The remaining eight proofs span sphere-packing density bounds (the first asymptotic improvement since 1978, reaching the Cohn-Elkies theoretical ceiling), three entries from Erdős’s catalogue of combinatorics problems including multicolor Ramsey numbers, circuit complexity lower bounds for computing the permanent, quantum parallel repetition, closest vector problem hardness, and Ehrhart’s volume conjecture. They are listed in the repository README with individual file names for each.

The compute cost for all ten successful proofs was roughly $2,000 at GPT-5.6 Sol API rates. DataCamp notes the critical qualifier: that figure reflects the tokens used in the runs that succeeded, not the total exploration budget. OpenAI has not disclosed how many failed attempts at these or other problems preceded the ten published solutions. $2,000 is the cost of the harvest; the cost of the search is elsewhere. OpenAI also attempted Millennium Prize Problems, including the Riemann hypothesis and P vs. NP. None of those succeeded. Astra is capable and bounded simultaneously, which is how useful tools generally work.

Astra itself is not a publicly available model. According to Byteiota’s coverage, it operates as a multi-agent system: a root agent coordinates subagents across extended timeframes — hours or days — distributing work, synthesizing results, and maintaining coherence across a task that exceeds any single context window. OpenAI demonstrated it privately to U.S. senators before the GitHub disclosure. Any public release will require clearance under the administration’s Covered Frontier Technology review framework, making Astra the first model that will need government approval before launch. The math papers are the announcement; the announcement is also lobbying.

What Lean Verification Doesn’t Give You

There is a careful distinction between what Lean verifies and what a Lean proof means. The kernel confirms that the formal proof is internally consistent — that every step follows correctly from the previous step under the rules of the type system. What it cannot confirm is that the formal statement in Lean correctly encodes the mathematical question the community has been asking. A Lean proof of a theorem is a proof of whatever the theorem says in Lean’s formal language. If the formalization is slightly wrong — if the encoding of “sofic group” or “von Neumann algebra” in Lean’s type system captures something adjacent to but distinct from what Gromov or Connes meant — the proof compiles and the open problem remains open. Lean removes one layer of required trust. It does not remove the need for mathematicians to read the formal statements and confirm they match the intended problems. That human verification step is still pending for all ten results.

The comprehension problem is separate and deeper. The 249-page manuscript contains both human-readable walkthroughs — the model’s mathematical reasoning, translated into prose — and the Lean certificates. Byteiota’s coverage describes Astra’s architecture as one where subagents synthesize mathematical arguments across long sessions. What the certificates produced by that process look like to a human mathematician is documented in a comment by software engineer Fernando Borretti in SiliconAngle’s coverage: “the frontier of the field will recede past the point where anyone can follow it.” This is not a critique of the proofs’ validity. It is a description of their shape. Lean proofs are machine-typed programs. A 200-line Lean certificate for a non-sofic group is correct by construction, and it is also — in any meaningful human sense — not a proof in the way mathematicians have historically used that word. It does not communicate why the thing is true. It certifies that it is.

DataCamp’s expert commentary puts this plainly: “A correct proof that nobody has fully audited or absorbed hasn’t really been understood yet, just verified.” This is a new category. Mathematics has always generated results before the community finished processing them — Ramanujan’s notebooks contained theorems Hardy spent years verifying. But those theorems were at least legible; a human could follow the argument even if they hadn’t yet confirmed each step. What AI-generated Lean proofs introduce is a form of mathematical knowledge that may be permanently out of reach of human comprehension. Not because mathematicians aren’t trying, but because the proof was generated by an optimization process that has no obligation to produce human-readable steps. Fields Medal winner Timothy Gowers endorsed OpenAI’s earlier unit distance conjecture work for publication in Annals of Mathematics, describing it as “beautiful.” Whether the non-sofic group construction earns comparable endorsement for the walkthroughs — as distinct from the certificates — is an open question of a different kind.

The community also needs to confront the velocity problem. Ten unrelated results across six mathematical subfields, in parallel, in a weekend. The review capacity for each one individually runs months to years. Thomas Bloom can tell you Erdős problem 183 is resolved because the certificate compiles. He cannot yet tell you whether the proof is correct in the sense mathematicians have always meant — whether someone followed the argument and found no errors. The Lean kernel has replaced that step for internal consistency. It has not replaced it for meaning.

Astra’s ten proofs represent the clearest demonstration yet that the trust problem in AI mathematics is solvable, and that solving it exposes a harder problem underneath. When you remove the need to trust the model or trust a human validator, you find that the thing you wanted — mathematical insight, not just mathematical truth — was never what the kernel was verifying. The non-sofic group exists. The Lean file compiles. No mathematician alive fully understands the proof that proves it. That is the state of things, and it is going to be the state of things for a while.

AI-generated editorial illustration · TemperatureZero · August 3, 2026