Post-Cutoff

Independent re-run of the Lean check for OpenAI’s ζ(s) ≠ 0 for Re(s) > 7/8 proof: “the proof checks”

Dave Goldblatt @davegoldblattGitHub

Why it matters

First public independent check of a headline claim: two kernels accept the Lean proof of the 7/8 zero-free half-plane.

Summary

Repository created 01:36 UTC Oct 7 (posted to HN as “Verified Riemann Zeta in Lean”). It re-runs the Lean check of OAI.riemannZeta_ne_zero_of_seven_eighths_lt_re from openai/math at commit adc7f12 (Lean/Mathlib v4.34.1). Lean FRO’s comparator with OpenAI’s challenge, and with a challenge written independently for the check plus the independent nanoda kernel (Rust), both accept. The only axioms are propext, Classical.choice and Quot.sound, and the statement uses Mathlib’s own riemannZeta. The proof is about 2,900 files and close to half a million lines. Caveats from the README: one operator on one machine, OpenAI’s repository assumed non-adversarial, and the written paper was not reviewed. Goldblatt’s X bio links him to Meta and two start-ups; he is not identified as a mathematician.

Archived text

README read on 2026-10-07 (raw GitHub). It also states that the check “was carried out by Claude Code, Anthropic’s coding agent, on the repository owner’s machine at the owner’s request”, and that GPT-6 Astra (via Codex) and Claude Fable reviewed the report, not the proof. Short quotes:

  • “Bottom line: unless both checkers are broken in the same way, the proof checks out.”
  • “Two independent kernel implementations, Lean’s own and nanoda (written in Rust), accept a proof of the statement from Mathlib’s definitions using only Lean’s three standard axioms.”
  • “the new content is the strip 7/8 < Re(s) < 1.”
  • “One operator ran both checks, on one machine. Nobody has yet reproduced it independently from a fresh clone.”
  • “OpenAI’s configs leave the second kernel off (402 of 405 set it to false). The second run turned it on.”
  • “This covers the Lean proof only, not the accompanying paper.”
Source: github.com/davegoldblatt/openai-zeta-proof-check

Cited in

  1. Science & math 98 days after the cutoff

    OpenAI releases 722 AI-written math manuscripts claiming hundreds of open problems

  2. Science & math 98 days after the cutoff

    OpenAI model claims the quasi-Riemann hypothesis