Work

Two AI proofs of critical percolation, checked against a statement written independently from the textbook

Year: 2026
Type: preprint
Abstract: In August and September 2026, Anthropic (J. Leder with Claude) and OpenAI released Lean 4 proofs that Bernoulli bond percolation on Z^d has no infinite cluster at the critical probability, θ(p_c) = 0, a question open in dimensions 3 to 10. The proofs are theirs. Before reading either, I wrote the statement in Lean from Grimmett's Percolation (2nd ed., §1.3–1.4), with Mathlib only and encoded differently from both (edges as unordered pairs at ℓ¹ distance one, the product measure by Measure.infinitePi, p_c as Grimmett's supremum), and checked, with Claude, that both proofs prove it: Anthropic's for every d ≥ 2 and OpenAI's for d = 3, each by a Lean bridge (299 and 306 lines; the monotonicity of θ in OpenAI's case proved by us). Both bridges are accepted by Lean's kernel and by nanoda with the independent statement as the Comparator challenge, and leanchecker --fresh replays both environments; Anthropic's development replays as shipped; planted false statements, one of them proved false in Lean in the same environment, are refused by both bridges; θ(0) = 0 and θ(1) = 1 are proved in the independent encoding, so the statement is not vacuous. What stays open: the two bridges run on two Lean versions, so that the statement means the same in both is a reading, not a kernel check; neither lab's informal paper is checked. Rules were sealed before each stage and results after, with a public registry and Bitcoin anchors. Files: the report (PDF) and the records, every file pinned by its hash.
Language: English
Subfield: Software
Open Access status: green
OpenAlex ID: W7225892823
Cites: 1
Cited by:
Related to: 10