AI methods and verification

1. 07

   Sep 30, 2026

   Papers accepted as posters at two NeurIPS 2026 workshops

   “Rediscover Before You Discover” was accepted at the MATH-AI workshop and “How Far a Wrong Claim Travels Before It Is Caught” at AI-Native Academia. The first asks an AI discovery pipeline to re-derive a recent human find before it may claim anything new; the second is a census of my own program’s errors, each with what caught it and how far it had travelled first. Both are free to read on Zenodo.

   Checked outside this site

   • WorkshopAccepted by the program chairs of MATH-AI, a NeurIPS 2026 workshop (a poster; the workshop is not archival), Sep 28, 2026. the paper on Zenodo
   • WorkshopAccepted by the program chairs of AI-Native Academia, a NeurIPS 2026 workshop (a poster; the workshop is not archival), Sep 30, 2026. the paper on Zenodo

   AI methods
2. 08

   Oct 20, 2026

   A paper on checking AI agents’ work, accepted at the Lifelong Agents workshop

   “Verify Before You Accept” studies how my team of AI agents checks work before accepting it. A run is one task a helper agent carried out and reported back to the lead agent, such as computing a number or checking a claim. Over the 33 days of records the paper covers, the lead agent’s check at acceptance changed 7 of the 202 runs it ruled on, and the incident records show at least 11^▲ of the 199 accepted runs were later found wrong. It was accepted as a poster at the 3rd Workshop on Lifelong Agents, part of AIAS+ 2026 in San Francisco, and the paper and its data are free on Zenodo.

   Checked outside this site

   • WorkshopAccepted by the program chairs of the Lifelong Agents workshop, AIAS+ 2026 (a poster; the workshop is not archival), Oct 20, 2026. the workshop
   • ArchiveCamera-ready paper and data posted, Nov 6, 2026. Zenodo DOI

   AI methods

Articles

• What a long-standing error in a math database means for anyone who copies reference valuesWriting · Oct 24, 2026A dropped digit sat in an OEIS entry for years until a Claude agent spotted it. What that means for anyone who copies values from reference tables.
• Verify Before You AcceptPaper · Oct 20, 2026A case study of my own AI agent team: what the lead agent's check before accepting work caught, and what got past it.
• Before my AI pipeline could hunt for new shapes, it had to rebuild a known oneWriting · Oct 4, 2026An entrance test for an AI-built search program caught a rule that would have thrown away the very solid it was looking for. Then a reviewer's pointer showed the "new" solid had been printed, unnamed, in a 2016 survey.
• A new best-known dominating set for the twelfth Fibonacci cube, and a month-long run with no answerWriting · Sep 26, 2026A network puzzle: pick as few points as possible so that every point is picked or linked to a picked one. A search I ran with Claude found a smaller set than any published.
• An OEIS editor asked what my "certificates" proved, and I changed how I label evidenceWriting · Sep 26, 2026When you hand someone results made with AI agents, they need to know what they can check from what you gave them and what they have to take on trust. An editor of a public math database asked me exactly that about a count I had added. The honest answer was "at least this many", not "exactly this many".
• The record I keep of errors my checks caught, and two that changed how I workWriting · Sep 11, 2026I keep a record of every error a check caught in my math work with AI agents: who made it and what caught it. Two of them changed how I check everything.
• How results here get checked, and what a full recheck after an AI model switch foundWriting · Jul 17, 2026When I rechecked everything an AI model produced during a switch I had not been told about, every computed number came out the same when it was recomputed. The errors were in the wording around the numbers and in how some checks had been done; one real finding had also gone unnoticed.
• Extending an expert's tables for unsolved take-away games in a way he could check himselfWriting · Jul 16, 2026Every run first had to reproduce the expert's published tables exactly. That let him check the new work on his own machines instead of taking my word for it.