On October 6, 2026, OpenAI released hundreds of mathematical manuscripts written by an unreleased model, claiming answers to long-standing problems. One of the mathematicians advising on the release described what happens next: no human understands such results when they are released, “and now the work begins.” A month later, that work is visible. A search of arXiv for papers whose abstracts respond to the release finds 102 in four weeks, about two dozen a week. Twenty-five rewrite a machine proof for human readers — simplifying it, re-proving it another way, or extending it to new settings. Seventy-four build on a result, seventeen of them explicitly assuming it is correct. Three question a result or the way it was checked. The uptake is uneven. The matrix-multiplication bound has drawn seven papers, including an independent constructive proof. The Harary–Hill crossing-number conjecture has two expositions. The Landau–Siegel result has been simplified, made explicit and extended. Several problems this library covered have not yet drawn a single paper. Twelve more papers show researchers using OpenAI and Anthropic models in their own discoveries. The headlines came on day one. Absorption is slower, uneven, and countable.
OpenAI’s release on October 6 put hundreds of model-written manuscripts on GitHub; the next day three were withdrawn over a sign error and fourteen revised, and by its own count about 42% of the headline results have Lean formalizations.[1] The advisory group of mathematicians that OpenAI consulted said the correctness of the collection had not been established, and Terence Tao reported having reviewed about ten of the papers.[16] Melanie Matchett Wood, a member of that group, put the situation plainly: no human understands such results when they are released, “and now the work begins.”[17]
This case counts that work. A search of arXiv on November 5 for papers since October 6 whose abstracts refer to the release, read and classified one by one, finds 102 human responses: 25 in the first week, 23 in the second, 24 in the third, 21 in the fourth, and 9 in the first three days of the fifth.[18] Twenty-five rewrite a proof: an exposition, a simplification, a different proof, or a check by an independent method. Seventy-four build on a result — extending it, applying it, or using it as one step in a new theorem; seventeen of those say outright that they assume it. Three question a result or how it was verified.
The rewriting is where belief is earned. For the bound of 9/4 on the matrix-multiplication exponent — a large step below the previous record of about 2.37 — Rafael Pass gave “a self-contained constructive proof,” Amanov, van den Berg and Zuiddam extended it to every field, and others followed.[2][3] For the Harary–Hill conjecture on the crossing number of the complete graph, open since the 1960s, Martin Balko and Gelasio Salazar each wrote an exposition, Salazar with an extension.[4][5] For the claim that Landau–Siegel zeros do not exist, Bryce Kerr published “a simpler proof” within ten days.[6]
The building is faster, and riskier. The physicist Hal Tasaki used OpenAI’s Haldane-gap proofs to prove a new uniqueness result.[7] Frei, Sofos and Widmer used the release’s zero-free result to make a well-known class-group bound unconditional.[14] Kozma derived a faster counting algorithm from the claimed refutation of the Exponential Time Hypothesis, and several others did the same.[10] Each of these inherits whatever the original proof gets wrong — the pattern this library first saw in UC-329, at a much larger scale.
25 rewrite a proof, 74 build on a result (17 explicitly conditional), 3 question one. A floor: from one search of abstracts on November 5, 2026.[18]
The questioning is small in number and large in substance. Bastounis, Circelli and Hansen argued, with the Navier–Stokes claim as their example, that a Lean check of an AI formalization “does not guarantee” the human-readable argument it accompanies.[12] Lekili and Sheridan examined the withdrawn Hodge-conjecture papers and the sign error behind them.[13] Dias and Lima set out the limits and constants of the ETH refutation.[11] At the same time, Zhen Lei and Xiao Ren began a two-part “readable and accessible account” of the Navier–Stokes blow-up — the same result, being both questioned and explained.[15]
The problems this library covered show the spread. The plane-colouring result of UC-328 is already a tool: its rigidity theorem, which has a machine-checked proof, is the key input to two new colouring papers.[9] The hot-spots result of UC-334 was extended by Krzysztof Burdzy within two days.[8] The ETH refutation of UC-339 is being built on before its review has reported.[10][11] The integer-multiplication claim of UC-329 appears only in conditional work. The Boltzmann, Thorp shuffle and Heilbronn results of UC-331, UC-332 and UC-333 have drawn no paper in this search yet.
Twelve further papers in the same weeks describe researchers using AI models directly — results whose main idea came from an OpenAI or Anthropic model, credited in the paper. One functional-analysis paper says its main idea “was found by Claude Fable 5” and documents how.[19] That is a second, quieter shift alongside the release: machines as named contributors inside human papers. The limits of this case: the count is a floor from one search method; classifying a paper by its abstract is a judgment; and being rewritten or used is evidence of engagement, not proof of correctness.
No human understands such results when they are released, and now the work begins. — Melanie Matchett Wood, as reported, October 2026 [17]
| Dimension | Evidence |
|---|---|
| Employee (D2) L1 · 88 | 102 human responses in four weeks: 25 rewriting, 74 building on, 3 questioning.[18]About Two Dozen Papers a Week |
| Quality (D5) Origin · 82 | About 42% formalized in Lean; three withdrawn and fourteen revised in the first day.[1]Some Checked, Some Withdrawn |
| Operational (D6) L1 · 76 | Public manuscripts and Lean files let human work start within days, on arXiv.[1][8]Open by Default |
| Regulatory (D4) L2 · 74 | 17 papers explicitly assume a result; a critique argues Lean does not certify the readable proof.[12][18]Building Before Checking |
| Customer (D1) L2 · 70 | A physicist and number theorists already use the results in new theorems.[7][14]Fields Start Depending |
The cascade originates in D5 (Quality): a large release of machine-written proofs, some formally checked, some withdrawn.[1] It propagates to D2 (Employee), the mathematicians who spend their weeks explaining, re-proving and using the results,[2][4][6] and to D6 (Operational), the open processes — GitHub, Lean, arXiv — that let that work start within days.[1][18] It lands in D4 (Regulatory), the norms of acceptance now being tested by conditional citations and formal-check critiques,[11][12] and in D1 (Customer), the fields that begin to depend on the results.[7][14] D3 (Revenue) is not scored.
It is the follow-up to the library’s mathematics cases. UC-328 asked whether specialists would read a machine’s proof; they did, and are now using it. UC-339 and UC-340 recorded claims waiting for review; one is being built on, one waits. The general shape: the answers arrived first, and understanding is arriving at the speed of people — about two dozen papers a week, concentrated on a few results. The review in January will repeat the search and see which results moved from built-on to rewritten, and which drew their first paper.
-- UC-342: Now the Work Begins: 6D Amplifying Cascade
-- OpenAI release Oct 6 2026 (openai/math): Oct 7 three withdrawn over a sign error, fourteen revised, about 42 pct of headline results formalized. Advisory group: correctness not established; Tao reviewed about ten. Wood: now the work begins. arXiv search Nov 5 2026, abstracts since Oct 6 referring to the release, classified by reading: 102 human responses; 25 rewrite, incl. one independent concurrent result (Pass constructive omega at most 9/4; Amanov, van den Berg, Zuiddam every field; Balko and Salazar Harary-Hill; Kerr simpler Landau-Siegel; Lei and Ren Navier-Stokes account); 74 build on, 17 explicitly conditional (Tasaki Haldane gap; Frei, Sofos, Widmer unconditional class groups; Kozma from ETH); 3 question (Bastounis, Circelli, Hansen on Lean; Lekili and Sheridan on the sign error; Dias and Lima on ETH constants). Weekly: 25, 23, 24, 21, then 9 in three days. 12 more papers credit OpenAI or Anthropic models with key ideas. Library: UC-328 used, UC-334 extended, UC-339 built on, UC-329 conditional only, UC-331, UC-332, UC-333 no papers yet.
FORAGE now_the_work_begins
WHERE large_ai_proof_release = true
AND human_responses_counted = true
AND results_used_before_review = true
ACROSS D5, D2, D6, D4, D1
DEPTH 3
SURFACE now_the_work_begins
DIVE INTO absorption_underway
WHEN rewrites_concentrated_on_few_results = true
AND most_results_unreviewed = true
TRACE reception_cascade
EMIT now_the_work_begins_signal
DRIFT now_the_work_begins
METHODOLOGY 88
PERFORMANCE 42
FETCH now_the_work_begins
THRESHOLD 1000
ON MONITOR CHIRP high 'One month after OpenAI released hundreds of machine-written proofs, a search of arXiv finds 102 human papers responding: 25 rewriting a proof, 74 building on a result, 17 of those explicitly conditional, and 3 questioning one. Uptake concentrates on a few results; several draw none yet.'
SURFACE analysis AS jsonRuntime: @stratiqx/cal-runtime · Spec: cal.semanticintent.dev · DOI: 10.5281/zenodo.18905193
About two dozen human papers a week, steady for four weeks. Reception can be measured, not just felt.[18]
An independent constructive re-proof of the matrix bound says more than any headline.[2][3]
Seventy-four papers build on results; seventeen say they assume them. Errors would travel.[18]
Primary sources are OpenAI’s release repository and the responding papers on arXiv, read at source. The count is this case’s own: an arXiv search on November 5, 2026 for abstracts since October 6 that refer to the release, each paper read and classified by hand. It is a floor, not a census. Source 20 is an interactive companion made for this case.
This case measures reception rather than announcing it: who explained, who built, who questioned. Organisations adopting AI-made answers can measure the same way. Book a discovery call to design the check.