Claude’s Percolation Directory Isn’t on Anthropic’s Main Branch
Open the repository on main and the directory is gone. A 28 August commit still has it, and nothing on record says why.

Open Anthropic’s public repository formal-math the ordinary way, on the branch named main, and the percolation directory is not there. On 3 October 2026 the root held a LICENSE file, a short README and one project, zeta23. Asking GitHub for percolation on that branch returns not found. The files have not disappeared from the repository. They are still in a commit from 28 August, if you already have the identifier.
It is the last of twelve commits that add the directory. The project’s notice says the Lean sources were written by Anthropic’s Claude models under the direction of Justin Leder, and that no human wrote or edited the Lean code. The README names Leder as the author. There is no launch note in the repository. Mathematicians mostly met the files through Gil Kalai, who on 3 September described a Claude document with a Lean check claiming a solution to a conjecture of Gady Kozma and Shahaf Nitzan, and wrote that he was not sure it was an official Claude release.
The claim, in plain terms: bond percolation opens or closes the links of a grid at random, and there is a critical probability. At that value, is the cluster of the origin almost surely finite? The answer was known for a flat grid and for very high dimensions, and open from dimension three through ten. The files claim a machine-checked argument that the cluster is finite from dimension two upward, by proving the Kozma–Nitzan conjecture meant to imply it. Their audit counts 251 Lean files and 97,574 lines, a successful build, and the comparator line “Your solution is okay!” Those figures were not re-run here. The README adds that a machine check does not decide whether the formal statement matches the intended theorem, and that no one independent of the author has refereed the work. That is context inside the files, not the oddity of the repository.
The files are still at one commit. 795efb86f191735c5481675763537cfb4ff37e55, made at 21:22 Rome time on 28 August 2026, contains the directory: the Lean tree, the statement file Challenge.lean, the audit, a fifteen-page guide. On that commit the root README lists percolation next to zeta23, and a CODEOWNERS file assigns the folder to the GitHub user jleder3. The project’s own links point at the same path on main. That path is the one that is missing. CODEOWNERS itself is absent from main.
Those twelve commits are not in the history of main. They start from 3d3e9f83, the merge of pull request 20, which is on main, and the first of them lands about an hour later, at 16:54 Rome time on 28 August. The histories then split. Main goes on to the merge of pull request 24 at 23:37 Rome time that evening, and a small tooling update on 5 September. GitHub reports the comparison as diverged: twelve commits on the percolation side, two on main. No current branch contains the directory. The commit has no pull request and no comments. Pull request 18, already merged, had said a percolation project would follow in a separate pull request; none of the repository’s pull requests carries these commits. Issues, pull requests and the README on main do not say why the directory is absent.
Checked again on 3 October 2026, that is as far as the public repository goes. The directory is not on main. The files remain in the commit, and GitHub still serves them. No source found there states a reason. It is also undocumented whether those commits were ever the tip of main and later dropped, or only ever stood on a line that never joined it. Nothing in the issues, the pull requests or the commit messages says the work was withdrawn, retracted or judged wrong. This note does not supply that conclusion.
A research artifact other people have been reading can remain in a public commit while the default view of the same repository no longer shows it, with no note on the difference. The project’s link to main fails. The commit identifier still opens the files. For work that is supposed to be checked by someone else, those are different offers.
percolation/ absent from main, still in commit 795efb86
On 3 October 2026 percolation/ was not on branch main of anthropics/formal-math (root lists only zeta23; the contents API returns 404). The files remain at commit 795efb86, 21:22 Rome time on 28 August 2026, on a 12-commit line that diverges from main at 3d3e9f83. No issue, pull request, README note or commit comment found that explains the absence or calls the work withdrawn. Lean was not rebuilt.
Open formal-math on main Open the 28 August commit that still has the directory See GitHub’s comparison: the histories have diverged Read Gil Kalai’s 3 September note, where many mathematicians first met the filesGitHub: anthropics/formal-math/tree/795efb86f191735c5481675763537cfb4ff37e55/percolation ↗
License: Apache-2.0
How this was written: researched and drafted with the help of AI tools, starting from the sources listed here. Editorial responsibility for what is published is mine. Corrections: hello@liminalfinds.com.