Interpreting claims that a theorem is correct because it was checked in Lean
- ID
- 47196
- Status
- summarized
- Published
- 16 Nov 2026, 3:38 PM
- Fetched
- 17 Nov 2026, 4:54 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://xenaproject.wordpress.com/2026/11/12/and-the-proof-is-checked-in-lean/
- Source URL
- https://hnrss.org/best
Summary
- Score
- 7.0
- Created
- 17 Nov 2026, 4:55 AM
- Tags
- Audience
- developersai_ml_learnersai_agent_usersvibe_coders
What happened
A Xena Project post lays out four questions to ask before accepting "this theorem is correct because it was checked in Lean," focused on AI-generated Lean code. It uses OpenAI's recent dump of 722 mathematics papers — about 45% shipped with Lean code — as the worked example, noting three of the non-Leanified papers were already retracted over a sign error, and that for many dumped papers the checks haven't been done. The core technical point: Lean rigorously separates a theorem's statement from its proof, so a headline statement can compile perfectly while being trivially false — the post's sample "FermatsLastTheorem" definition compiles but fails because Lean's ℕ includes 0, giving a=b=c=0, n=3 as a counterexample.
Why it matters
The failure mode is the one builders hit daily with AI-generated artifacts: a green check proves the artifact matches the spec you wrote, not that the spec matches what you meant. The Fermat snippet is the concrete case — it type-checks, yet is false — so "it compiles / the test passes / the agent verified it" carries no weight until someone confirms the statement itself. If you accept AI-produced code, proofs, or config with a passing check as the correctness claim, you're skipping the step the post says was skipped for many of the 722 papers; the retractions happened in the subset without formal code, which is exactly the asymmetry to watch.
Discussion angle
What is your team's version of Question 1 — who reviews the spec or statement an AI wrote, rather than the proof or test that it passes — and would the 722-paper dump's 45% Lean-coverage rate have made you trust it more or less?