Reader finds 10+ results in OpenAI's math repo whose headline claims fall outside their Lean checks
MysteriousAvocado580 · reddit · 2026-10-11
A Reddit user read all 242 Lean scope notes in OpenAI's math repo (openai/math, commit fd4aeeb, Oct 8) and found that in at least 10 result families, the note explicitly states the headline result is not what the Lean formalization covers — yet the headline still carries a "(Lean)" tag.
Key examples:
- #266 "Exactly three mutually unbiased bases in dimension six": Lean only proves a weaker bound (families of at most five members)
- #195 "Counterexample to the small Cohen–Macaulay module conjecture": only the candidate ring is verified
- #312 "Grothendieck homotopy hypothesis": only the elementary-expansion theorem is formalized
- #084 "Geometric case of the Erdős similarity conjecture": headline claims all q in (0,1), Lean covers q = 1/2
- Other affected families include entropy thresholds, reconstruction thresholds, the separable quotient problem, and quantum secret key distillation
The author stresses this goes beyond the general "Lean checks a statement, not the claim" critique: OpenAI's own scope notes admit the headline conclusions are outside the formalized statements. Four of the ten (027, 066, 195, 312) also have papers affected.
More from Research
- Anders Sandberg's Claude-made video shines with Tromp's Lambda Diagrams notation — anderssandberg · 2026-10-11
- Lean Co-creator Jeremy Avigad on AI, Verification, and the Future of Mathematics — EchoShao8899 · 2026-10-11
- NeurIPS 2026 position paper argues LLMs lack explicit cognitive control for true autonomy — AndrewLampinen · 2026-10-11
- PENDULUM: A Friction-Free Open-Source Interpretability Tool Built on the Jacobian Lens — Extraaltodeus · 2026-10-11
- LeaP from Southeast University, accepted at CoRL 2026, makes diffusion action priors state-aware — jiqizhixin · 2026-10-11
- Build your own decision model: a hands-on tutorial without the black box — softwaredoug · 2026-10-11