Crack-tip has no calibration (GPT-6 Astra idea)
Event confirmedAwaiting review
The takeaway
arXiv 2611.26245 shows that the convex lifting of the Mumford–Shah functional by calibrations (Alberti–Bouchitté–Dal Maso) is not sharp in the plane: for crack-tip boundary data a lifted competitor has strictly lower energy, so the crack-tip admits no calibration. The authors say GPT-6 Astra suggested looking for a gap and proposed the initial competitor in September; the paper was then written without AI. One author, who spent years of a PhD trying to calibrate the crack-tip, writes that knowing the key idea came from an AI ‘leaves me with this sense of bitterness’.
Status
- Claim
Event confirmedAwaiting review
- Our reporting
- High confidence
- Verification
- Preprint only
- Importance
- 3 of 5
- Last verified
- 20 November 2026
Your AI and this story
- GPT-6.2 Astra172 days after its cutoff
- Claude Opus 5.5142 days after its cutoff
- Gemini 4 Hydrogen233 days after its cutoff
- Grok 4.7172 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 142 days before it.
Key facts
- Result (abstract): ‘the convex lifting of the Mumford-Shah functional constructed through the notion of calibration introduced in [2] is not sharp in dimension two … As a consequence we establish that the crack-tip admits no calibration’
- AI statement (verbatim): ‘Following several unsuccessful attempts to construct a calibration for the crack-tip, GPT-6 Astra suggested investigating the existence of a positive gap in the convex lifting and subsequently proposed the initial competitor. The paper is built on the initial idea and it has been integrally written without the use of AI, apart from a grammar and a notational check’
- Author’s reflection (verbatim): ‘Having spent years of the PhD trying to calibrate the crack-tip … finally having an answer to this question feels bittersweet almost ten years after completing my PhD … knowing that the key idea came from an AI leaves me with this sense of bitterness’
- Context: calibrations are the standard route to proving that a candidate is a Mumford–Shah minimizer; the result rules this route out for the crack-tip in the convex-lifting framework. An unchecked ChatGPT Astra proof of the planar Mumford–Shah conjecture was posted in September (2026-09-22 entry)
- Status: preprint; no reaction found by 20 Nov
What happened
On 19 November 2026 Carioni and Patrini showed that the convex relaxation of the Mumford–Shah functional built from calibrations has a strictly positive gap in the plane, using a competitor first proposed by GPT-6 Astra in September after their own attempts to calibrate the crack-tip failed. The paper closes with a personal note in which the author describes the result as bittersweet because the key idea came from an AI.
Why it matters
It answers a question that had resisted specialists for years, and the authors draw a clear line between the AI’s contribution (the idea and the first competitor) and theirs (the proof and the writing). The author’s note is an unusually direct record of how a researcher experiences such a result.
What is disputed or not yet verified
| Verification | preprint, not refereed |
|---|
Sources
1 source from 1 site. Numbers match the chips in the text.
1 source: 1 primary
Primary
Changes
- Filed (science agent; PDF read)