A non-human mind produces knowledge beyond us

An artificial system originates a scientific or mathematical idea, not merely a search that exhausts the cases: the key construction or proof idea comes from the system and not from its operators, and the record says so. A brute-force proof that a solver ground out is a tool's work; a construction nobody had thought of is a mind's. I had to decide how strictly to read "not from its operators", and the answer I came to is this: when the operators built the search frame and the scoring rule, then the idea is theirs, even if the system found the answer. On that reading the machine-found constructions of 2022 to 2025 that beat the best human ones do not meet it: not FunSearch, not AlphaTensor, and not AlphaEvolve for now, since no peer-reviewed record yet says its key idea was the system's. Until October 6th, 2026 this page read the test more loosely and counted those results as plausibly meeting it.

The test: what counts as it having happened

A result in mathematics or the natural sciences whose key construction or proof idea was generated by a learning-based or search-based artificial intelligence system and not specified by its operators, acknowledged as such in the peer-reviewed record, that is verified by machine checking or by experiment and is acknowledged in that record as beyond what human researchers had produced. An exhaustive case check or a satisfiability proof whose idea was the operators' does not count.

Settled by: The peer-reviewed record and the verification artefact

The odds, by how long I live

baselineto 2071 ± 8	moderateto 2080	strongto 2095	radicalto 2136	openno deadline
95%	96%	97%	98%	98%

Each column is a different guess at how long I live, from current trends (baseline) to open-ended. The number is the chance this happens before that deadline, assuming everything it depends on happens first. As of September 20th, 2026.

Why these numbers: Re-estimated September 20th, 2026 against the machine-originated-idea wording: at least two published results from 2023 to 2025 (a search-based system's new constructions for a long-open combinatorial problem, and a later system's improved matrix-multiplication and packing bounds) appear to meet it on their face, and the record in each case says the construction came from the system, so the remaining doubt is whether a scan finds the wording met in so many words. Proposed by Claude September 20th, 2026, I accepted. Note added October 6th, 2026: this reason was written before the strict reading settled the same day, under which those results do not meet the criterion (docs/definitions.md, decision log). The numbers are held as they are until the review after December 31st, 2026. Proposed by Claude, approved by me.

What this one is worth

Nothing. The headline number does not rest on this step at any remove, so however it turns out, the chance that I live and work off Earth does not change. That is not a judgement that it does not matter. It is what the tree says: this step feeds a tier above the one the headline asks about, or sits beside the equation rather than in it.

What the ledger says about this step

• OpenAI posted hundreds of machine-generated maths results on October 6th, 2026, but none is peer reviewed yet, so the first Contact Clause rung stays where it was 2026-10-11 Widely reported, moved nothing · this step was read against it and did not move · found by the weekly web scan
• The first rung of the Contact Clause was rewritten to ask for a machine-originated idea, because as written it had already been met by a brute-force proof in 2016 2026-09-20 The tree itself changed

The last 3 checks

Once a week the tree is read against the news, node by node, against the test above. Open a line to see what that week held. Every check ever made on this node is on the Radar page.

• 2026-11-01 No change

  Verify claims, no change: AlphaEvolve mathematics paper is still arXiv 2511.02864 with no journal acceptance found; AlphaTensor facts hold (Nature, October 2022, 47 multiplications mod 2); the Erdős unit-distance disproof is still preprints only (arXiv 2605.20695 and an OpenAI note) with no journal record found. Separately, a secondary report says OpenAI announced on August 1st, 2026 an internal model 'Astra' that made progress on 10 open problems with Lean-formalized proofs, outside the window and not a peer-reviewed record.

  source 1source 2source 3source 4

• 2026-10-25 Ruled

  Verify claims: AlphaEvolve's mathematics paper is still the arXiv preprint 2511.02864 (November 3rd, 2025) with no journal acceptance found; the OpenAI Erdős unit-distance disproof (announced May 20th, 2026) has arXiv papers and outside checking but no journal publication found; AlphaTensor's Nature paper of October 5th, 2022 is unchanged and stays unresolved under the strict reading. New this week: Google DeepMind's AlphaProof Nexus is reported as published in Science on October 8th, 2026 (nine of 353 open Erdős problems and 44 of 492 open OEIS conjectures proved, every proof machine-checked in Lean), a peer-reviewed record the node's list does not yet name. I could not open the Science paper (science.org returned 403) or the arXiv preprint, so I cannot say whether it credits the key ideas to the system or calls the results beyond human work.

  For me: No move made. Proposal: add AlphaProof Nexus (Science, October 8th, 2026; arXiv 2605.22763; Lean proofs at github.com/google-deepmind/alphaproof-nexus-results) to C1's verify list as a candidate, to be read against the criterion by the next scan. The criterion needs the peer-reviewed record to say the key idea was the system's and that the result goes beyond what humans had produced; secondary coverage describes Gemini devising the proof strategy and Lean checking it, and says the nine Erdős solutions are not all full solutions of the original problems (Erdős 12 is only partly solved), so it may well not meet the bar. I could not fetch the primary pages (science.org 403; arXiv not fetched), so I am not proposing a correction, only that the paper be read directly.

  Taken off my list on November 1st, 2026, by a later scan.

  source 1source 2source 3

• 2026-10-18 Nothing new

  Verify claims: AlphaEvolve's mathematics paper is still the arXiv preprint 2511.02864 (submitted November 3rd, 2025) with no journal acceptance found, so no peer-reviewed record yet says the construction was the system's. The OpenAI Erdős unit-distance disproof (announced May 20th, 2026) has arXiv papers and outside checking but no journal publication found. AlphaTensor's Nature paper of October 5th, 2022 is unchanged (47 multiplications for four-by-four matrices modulo 2) and stays unresolved under the strict reading. Also seen: Axiom Math says five journals have accepted proofs its prover made (Axios, May 26th, 2026, company claim, page could not be fetched) and OpenAI's Astra results (August 2nd, 2026) are unrefereed; neither shows a peer-reviewed record crediting the key idea to the system as beyond human work.

Revisions

Date	Field	From	To	Why
2026-09-19	resolution	any result produced by an artificial system, machine-verified, acknowledged as beyond human researchers, and not checkable by hand (structural: the criterion changed after estimation)	the result's key construction or proof idea was generated by a learning- or search-based AI system and not specified by its operators, acknowledged as such in the record; exhaustive case checks and operator-designed satisfiability proofs excluded	Step 27 outside-model review, item 90 (docs/reviews/2026-09-19-rulings.md): on the plain reading the node had already resolved on a 2016 machine-produced satisfiability proof of 200 terabytes, which is a search program's work, not a mind's; I chose option (b), reword to the machine-originated-idea form, keep open, re-estimate after checking the 2023 to 2025 candidates, logged as structure. Decided September 19th, 2026.
2026-09-20	probability	0.92 / 0.95 / 0.97 / 0.98 / 0.98	0.95 / 0.96 / 0.97 / 0.98 / 0.98	Step 27 outside-model review, item 90, the re-estimate owed after the reword: the 2023 to 2025 candidates appear to meet the new wording, so the near-window columns rise slightly; a scan resolves it or does not. Proposed September 20th, 2026, I accepted.
2026-10-06	description	Plausibly already resolved on machine-found mathematical constructions from 2023 to 2025 that beat the best human ones; the weekly scan checks those against the wording before anything is settled.	Read strictly, as ruled 2026-10-06, the machine-found constructions of 2022 to 2025 do not meet it: not FunSearch, not AlphaTensor, and not AlphaEvolve for now.	Decided October 6th, 2026 on the open questions left by the first scan audit, after the strict reading of clause (a) the same day (docs/definitions.md, decision log); proposed by Claude, approved by me. Words only: criterion and probabilities unchanged.
2026-10-06	rationale	(no note)	A dated note: the reason predates the strict reading of 2026-10-06, and the numbers are held until the review after 2026-12-31. The original reason is kept.	Decided October 6th, 2026 on the open questions left by the first scan audit, after the strict reading of clause (a) the same day (docs/definitions.md, decision log); proposed by Claude, approved by me. Words only: criterion and probabilities unchanged.
2026-10-06	verify	Candidates for resolution under the 2026-09-19 wording, written from memory on 2026-09-20 and to be checked at the next scan against the published record: FunSearch (Nature, December 2023), new constructions for the cap set problem found by a language-model-guided search and described by the authors as beyond the best known; and AlphaEvolve (May 2025), an improved algorithm for multiplying four-by-four complex matrices and improved bounds on several open problems including the kissing number in eleven dimensions. Neither is resolved here: under the Contact Clause null rule (docs/proposal-contact-clause-null-rule-2026-09-19.md) the criteria were written before the evidence was read against them, and the scan does the reading.	FunSearch marked settled by the 2026-10-06 ruling and not checked again until the review after 2026-12-31; AlphaEvolve kept open to check, narrowed to whether a peer-reviewed paper says its key construction came from the system and not from its operators.	Decided October 6th, 2026 on the open questions left by the first scan audit, after the strict reading of clause (a) the same day (docs/definitions.md, decision log); proposed by Claude, approved by me. Words only: criterion and probabilities unchanged.
2026-10-12	verify	Two claims to check: AlphaEvolve, narrowed to the peer-review question, and AlphaTensor, awaiting a check of its facts.	The AlphaEvolve claim carries the scan's finding of 2026-10-08 (preprint only, no journal acceptance found) and stays open; the AlphaTensor claim carries the same scan's confirmation of its facts; a third claim is added, the May 2026 OpenAI disproof of Erdős's unit-distance conjecture, to be checked for a peer-reviewed record.	The weekly scan of October 8th, 2026 read the list as the procedure requires and flagged both findings and the new candidate for me, since it may not edit the list (data/scans/2026-10-08.toml). Proposed by Claude October 12th, 2026, I approved at the merge. Words only: criterion and probabilities unchanged, and the strict reading of October 6th, 2026 stands.
2026-10-21	description	Read strictly, as I ruled on 2026-10-06: when the operators built the search frame and the scoring rule, the idea is theirs, even if the system found the answer.	I had to decide how strictly to read 'not from its operators', and the answer I came to is this: when the operators built the search frame and the scoring rule, then the idea is theirs, even if the system found the answer. [...] Until October 6th, 2026 this page read the test more loosely and counted those results as plausibly meeting it.	Two site-wide rules of October 21st, 2026: the prose a visitor reads says what was decided and why, in plain words, and spells dates out. The decision itself is unchanged, and its wording is kept word for word because an essay quotes it. Words only: criterion and probabilities unchanged.
2026-10-21	rationale	the strict reading I ruled on 2026-10-06	the strict reading settled the same day	The same two rules of October 21st, 2026. Words only: criterion and probabilities unchanged.
2026-10-21	verify	four phrases that described the strict reading as something I handed down	the same claims, saying plainly that it was my decision	The same two rules of October 21st, 2026. Words only: the claims themselves and what each scan checks are unchanged.

---

Think this estimate is wrong? Here's how to tell me, and what happens next. The record itself is data/C-contact.toml in the public repository.