OpenAI Ten Proofs
Ten supplied mathematical proof files were checked with Lean, software that verifies formal reasoning. Here is what passed, and why that is different from independently re-proving every claim in the accompanying written explanations.
Ten of ten source modules rebuilt. Twelve of twelve headline declarations accepted by the Lean kernel.
What was checked?
The OpenAI Ten Proofs project includes mathematical results expressed in Lean, a proof assistant. In Lean, a mathematical statement and its proof are written in a form that a small trusted checker can examine.
GCL independently rebuilt all ten supplied proof files and confirmed that twelve principal declarations were accepted by the Lean checker. This is stronger than merely running examples or observing that an argument looks plausible: the formal system checked the statements against their declared assumptions and dependencies.
It is also narrower than saying that GCL discovered ten new theorems or reverified every sentence of the project's written explanation. The check applies to the exact formal statements in a recorded version of the software and mathematical libraries. Substantially different wording or stronger claims require their own evidence.
Why keep the exact record?
Software and mathematical libraries can change. Recording the source version, Lean version and dependencies lets another researcher repeat the check on the same material. The technical tables below preserve that information for specialists.
The exact subject
| Field | Bound value |
|---|---|
| Repository | openai/ten-proofs |
| Commit | 94bc0feb6a9ff12c7d31d6de640a725c9d43d2b6 |
| Tree | 174289e4d4958cb0509874e6e53400e098213de7 |
| Lean | 4.32.0 |
| mathlib | 81a5d257c8e410db227a6665ed08f64fea08e997 |
| MATHCERT protected merge | 2aeb5875532152310331662e18327cf9dc3c736e |
Ten advertised results
| # | Result | Lean module | Disposition |
|---|---|---|---|
| 1 | High-dimensional sphere packing | SpherePacking |
Kernel verified |
| 2 | Binary and spherical codes | MetricCodes |
Kernel verified |
| 3 | Non-sofic groups | NonSoficGroup |
Kernel verified |
| 4 | Connes rigidity counterexample | ConnesRigidity |
Kernel verified |
| 5 | Permanent circuit and formula lower bounds | Permanent |
Kernel verified |
| 6 | Quantum parallel repetition | QuantumParallelRepetition |
Kernel verified |
| 7 | Closest-vector and decoding hardness | GapCVP |
Kernel verified |
| 8 | Sharp Ehrhart volume inequality | EhrhartVolumeInequality |
Kernel verified |
| 9 | Multicolor triangle Ramsey numbers | MulticolorTriangleRamsey |
Kernel verified |
| 10 | Extremal graph counterexamples | CompactnessAndDegeneracy |
Kernel verified |
What “verified” means here
Lean kernel acceptance of each headline declaration checks its complete formal
dependency graph. The axiom audit found only the programme’s permitted standard
axioms: Classical.choice, Quot.sound, and propext.
The final candidate received fresh independent non-author review on its exact commit before expected-head merge. The complete corpus workflow then passed again from protected main.
Evidence you can inspect
| Surface | Record |
|---|---|
| Machine-readable disposition | Protected corpus verification record |
| Exact integration and review | MATHCERT pull request #290 |
| Protected-main replay | Workflow run 34838818609 |
| Corrective tracker and closure | MATHCERT issue #289 |
| Upstream source tree | openai/ten-proofs at the verified commit |
The boundary remains visible
The exact supplied Lean formalizations for all ten advertised results were independently rebuilt and accepted by the Lean kernel on their recorded headline declarations, under the statement qualifications retained in the protected MATHCERT certificates.
This does not assert literal line-by-line identity with the accompanying PDF exposition. It makes no novelty, priority, authorship, or publication claim. It grants no authority to unlisted statements or broader paraphrases.
That precision is the point: GCL publishes what crossed the proof boundary—and keeps everything else on the correct side of it.