Vaught’s conjecture fails for infinitary logic
Ackerman, Freer and Knight post a Lean-verified counterexample, built over 8 months with Claude Code and Codex
Event confirmedAwaiting review
Importance: major (4 of 5)The takeaway
A sentence of L_ω1,ω with exactly ℵ1 countable models and no perfect set of them refutes the infinitary Vaught conjecture (1961). It follows Robin Knight’s 2002 approach, which was never accepted, and the whole argument is now certified in Lean.
Status
- Claim
Event confirmedAwaiting review
- Our reporting
- High confidence
- Verification
- Lean-verified
- Importance
- Major (4 of 5)
- Last verified
- 14 October 2026
Your AI and this story
- GPT-6 Astra163 days after its cutoff
- Claude Opus 5.5102 days after its cutoff
- Gemini 4 Argon193 days after its cutoff
- Grok 4.7132 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 102 days before it.
Key facts
- Paper: ‘A counterexample to Vaught’s Conjecture for L_ω1,ω’, Nathanael L. Ackerman, Cameron E. Freer, Robin W. Knight, arXiv 2610.15217 [math.LO], 26 pages, submitted 10 Oct 2026
- Result: a sentence of L_ω1,ω with exactly ℵ1 countable models up to isomorphism and no perfect set of non-isomorphic models, so the number of countable models is neither countable nor 2^ℵ0 in the strong (perfect-set) sense
- Scope: the paper disproves Vaught’s conjecture ‘for infinitary logic’; it does not claim the original first-order case (Knight’s 2002 manuscript claimed a first-order counterexample)
- History (paper, §1.2): Knight announced a first-order counterexample in 2002; despite a Notre Dame conference devoted to it, the argument ‘did not achieve general acceptance as a correct proof, although no definitive flaw in the construction had been identified’
- AI use: in Feb 2026 Ackerman and Freer proposed ‘to use AI coding agents to help formalize his proof in Lean’. ‘Over the following 8 months, Freer worked extensively with coding agents (Claude Code and Codex with models ranging from Opus 4.5 through 5.5, Fable 5 and 5.1, and GPT 5.2 through 6) and occasional help from GPT Pro (5.2 through 6)’
- Not a routine formalization: Knight’s object needed ‘substantial revisions’ to satisfy his conditions; ‘Closing this gap involved a great deal of human mathematical direction and AI-assisted exploration’. Finding the template system ‘required a great deal of work using coding agents to interact with LLMs, Lean, and SMT and Horn clause solvers’
- Verification: ‘a formally verified counterexample … with the entire argument certified in Lean’; one version checked by Comparator and registered as PALOMAR-
2026-10-07-000001; a cleaner version on GitHub (cameronfreer/vaught-conjecture-counterexample, created 30 Sep 2026); the paper notes that ‘not every line of this exposition has been formally verified’ - Tools: Freer’s lean4-skills pack for AI coding agents (Oct 2025), the Lean LSP MCP server, and Freer’s infinitary-logic Lean library
What happened
On 10 Oct 2026 Nathanael Ackerman, Cameron Freer and Robin Knight submitted arXiv 2610.152171 (listed 13 Oct). It gives a sentence of the infinitary logic L_ω1,ω with exactly ℵ1 countable models up to isomorphism and no perfect set of non-isomorphic models. Vaught’s conjecture (1961) says this cannot happen, so the paper disproves its infinitary version. The original first-order conjecture is not claimed.
The paper’s history section explains how it came about. In 2002 Knight announced a first-order counterexample. The community tried hard to check it, including a conference at Notre Dame, but the proof “did not achieve general acceptance”, although nobody found a definitive flaw. In February 2026 Ackerman and Freer suggested formalizing Knight’s newer, clearer infinitary draft in Lean with AI coding agents. Freer then worked for 8 months with Claude Code and Codex (Opus 4.5 to 5.5, Fable 5 and 5.1, GPT 5.2 to 6, plus occasional GPT Pro). The draft’s combinatorial object turned out to need “substantial revisions”; closing the gap took “a great deal of human mathematical direction and AI-assisted exploration”, and SMT and Horn-clause solvers were used near the end. The template system in the paper is a simplified version of the first one found.
The formal proof is on the Palomar Registry2 (Comparator-checked) and in a GitHub repository3 with a cleaner API. The authors note that the written exposition and the Lean development are organized differently in places.
No expert reactions or press coverage were found on 14 Oct (web search; X not searched directly).
Why it matters
Vaught’s conjecture is one of the best-known open problems in model theory. Here the human idea is two decades old; what changed is that coding agents made a full machine check practical, and the check forced repairs that the informal argument had missed. It is a clear case of AI-assisted formalization settling a disputed human proof rather than finding a new one. Whether the method extends to the first-order case is open.
What is disputed or not yet verified
| Verification | Formal proof in Lean (Comparator-checked, Palomar Registry); unrefereed preprint |
|---|
Sources
5 sources from 4 sites. Numbers match the chips in the text.
5 sources: 4 primary, 1 press
Primary
- arXiv 2610.15217: A counterexample to Vaught’s Conjecture for L_ω1,ωarxiv.org, paper
- Palomar Registry entry PALOMAR-2026-10-07-000001palomar-registry.org, code
- GitHub: cameronfreer/vaught-conjecture-counterexample (Lean 4)github.com, code
- GitHub: cameronfreer/lean4-skillsgithub.com, code
Press
- Wikipedia: Vaught conjecture (background, Knight’s claim)en.wikipedia.org, press
Changes
- Filed from the arXiv paper