AI systems have produced correct, novel, research-level mathematics, including the disproof of a 1946 Erdős conjecture in discrete geometry and medal-level performance at the International Mathematical Olympiad.

Unusually well anchored for a recent claim because the mathematics is independently checkable: the Erdős disproof is attested by an unaffiliated mathematician who builds on it, and the IMO results are Lean-verified and officially graded. The documented false-positive history in this area is a reason to anchor on primary mathematics rather than announcements, which this case does — not a reason to doubt the verified instances. Diagnosticity for the headline thesis remains low by scoping.

The AI disproof of the Erdős conjecture is verified by independent mathematicians, and AlphaProof's IMO performance was machine-checked in Lean and peer-reviewed in Nature. The historical false-positive rate for AI mathematical claims is heavily mitigated here by external evaluation. However, these milestones only bear on the philosophical motivation for Orch OR, entirely leaving the biological and physical mechanisms untouched.

Sawin’s arXiv paper independently states and builds on an OpenAI model’s disproof of a 1946 Erdős unit-distance conjecture; AlphaProof’s Nature paper documents Lean-verified IMO-level problem-solving; and the DeepMind announcement describes an officially-graded IMO gold medal under contest conditions in natural language. These sources together robustly establish that AI systems have produced correct, novel, research-level mathematics and competition-level performance. There is a documented history of earlier overclaimed AI–Erdős “solutions” collapsing upon literature checking, but the unit-distance result and IMO medals have cleared those hurdles. The main objection concerns interpretation, not fact: critics argue these are novel results within existing frameworks (deduction/induction), not the kind of abductive framework-generation Penrose’s argument targets.

The mathematics is independently checkable. Sawin’s arXiv paper states that an OpenAI team disproved Erdős’s unit-distance conjecture and improves the exponent to n^1.014; outside mathematicians checked the original construction. AlphaProof’s IMO-2024 solutions were Lean-verified and the methodology peer-reviewed in Nature; Gemini Deep Think’s 2025 gold-standard 35/42 was officially graded by IMO coordinators under contest conditions. A documented false-positive history on other AI-Erdős claims (October 2025 literature collisions; a Christmas 2025 withdrawal after a 1977 Erdős resolution surfaced) requires checking novelty, and does not touch these verified results. The strongest opposing consideration is conceptual rather than factual: competition problems and a novel construction inside an existing framework are, on Zahavy’s own taxonomy, mechanized deduction plus search, not the abductive framework-invention the Gödelian argument says is at issue.

Unusually well anchored for a claim this recent because the mathematics is independently checkable: the Erdős unit-distance disproof is attested by an independent working mathematician who states it in his own arXiv paper, names the verifiers, and sharpens the bound; AlphaProof's 2024 IMO performance was machine-checked in Lean with methodology peer-reviewed in Nature; and the 2025 gold was graded by IMO coordinators under contest conditions. The strongest opposing consideration is provenance and history: key elements rest on corporate self-report about unreleased models, the natural-language 2025 solutions were not machine-verified, and this evidence line has a documented false-positive history (October 2025 and Christmas 2025 retractions). Those caveats discipline how new claims are admitted — they do not dislodge the independently verified results on this record.

This is unusually well anchored for a claim this recent, because the mathematics is independently checkable and the case does not rest it on any company's announcement. The Erdős result is attested by an unaffiliated working mathematician who states it in his own arXiv paper, builds on the construction, and sharpens the exponent, with named outside mathematicians having verified correctness. The IMO evidence is firmer still: proofs machine-checked in Lean with methodology peer-reviewed in Nature, and a subsequent gold-standard performance graded by IMO coordinators under contest conditions. The documented false-positive history in this evidence line is a genuine caveat, but its failure mode was inadequate literature search rather than incorrect mathematics, and it does not touch the independently checked unit-distance result. The strongest opposing consideration is not about the facts but their reach: competition problems are solvable within an existing framework by construction, and the claim's relevance to this case runs only through ORCH-C040, the motivation rung, leaving every microtubule, anesthesia, and collapse claim exactly where it stood.

Independently checkable mathematics: the Erdős unit-distance disproof is attested and improved upon by an unaffiliated mathematician (ORCH-E022), and the IMO results are Lean-verified or officially graded (ORCH-E023, ORCH-E024). The documented false-positive history (ORCH-E027) argues for anchoring on primary mathematics, which this case does; diagnosticity for the headline thesis remains low by scoping.

Independently checkable mathematics: the Erdős disproof is attested and extended by an unaffiliated mathematician (ORCH-E022), and the IMO results are Lean-verified or officially graded (ORCH-E023, ORCH-E024). The documented false-positive history (ORCH-E027) argues for anchoring on primary mathematics, which this case does; diagnosticity for the headline thesis remains low by scoping.

Independently checkable mathematics: the Erdős disproof is attested and extended by an unaffiliated mathematician (ORCH-E022), and the IMO results are Lean-verified or officially graded (ORCH-E023, ORCH-E024). The documented false-positive history (ORCH-E027) concerns novelty checking, not correctness, and this case anchors on primary mathematics accordingly.

AI achievements in mathematics are independently verified. An AI model disproved an Erdős conjecture (confirmed by an independent mathematician building on the construction), and AlphaProof reached officially-graded IMO gold-medal standards with machine-checked Lean proofs. While there is a history of false positives in AI math claims, these specific milestones are robust. However, they are highly low-diagnosticity for the mechanism of Orch OR, bearing only on its historical philosophical motivation.

Sawin’s arXiv paper independently states that an OpenAI model disproved Erdős’s unit-distance conjecture and builds on that result, while Nature’s AlphaProof paper and the DeepMind IMO announcement document Lean-verified silver-medal and officially-graded gold-medal IMO performances, respectively. These sources collectively establish that AI systems have produced correct, novel, research-level mathematics and achieved medal-level contest performance. The strongest counterpoint is not factual but interpretive: some argue these feats remain within established frameworks, and that the systems’ novelty may ultimately be viewed as sophisticated interpolation over the human mathematical corpus, but this does not change the truth of the claim as worded.

The Erdős unit-distance disproof is attested by an independent mathematician's own arXiv paper that states the result, names the checkers, and sharpens the bound to n^1.014; AlphaProof's IMO-level solutions were Lean-verified and the methodology peer-reviewed in Nature; the 2025 gold-standard performance was officially graded by IMO coordinators under contest conditions. Those are checkable mathematical facts, not company announcements. The strongest opposing consideration is conceptual rather than factual: producing correct theorems inside existing frameworks is not framework extension, a DeepMind researcher who co-authored AlphaProof argues the systems remain structurally incapable of abductive premise-formation, and this evidence line has a documented false-positive history of "novel" solutions that were already in the literature.

Unusually well anchored for a claim this recent because the mathematics is independently checkable: the Erdős disproof is attested by Sawin's own arXiv paper stating the result and building on the construction rather than by OpenAI's announcement; AlphaProof's solutions were machine-checked in Lean with methodology peer-reviewed in Nature; and the 2025 IMO gold-standard performance was officially graded by IMO coordinators under contest conditions. The strongest opposing consideration concerns provenance and history, not the results: key elements rest on corporate self-report about unreleased internal models, and this evidence line has a documented false-positive history (October 2025 and Christmas 2025 claims collapsed on prior literature). Neither touches the verified results. The deeper objection — performance is not understanding — bears on what these results imply for ORCH-C040, not on the truth of this claim as worded, which is diagnostic only for the motivation rung and silent on microtubules, anesthesia, and collapse.

Unusually well anchored: AlphaProof's IMO solutions were machine-checked in Lean with methodology peer-reviewed in Nature, the 2025 gold-standard performance was officially graded by IMO coordinators under contest conditions, and the Erdős unit-distance disproof is attested by an independent working mathematician who states it in his own arXiv paper, names the verifiers, and sharpens the exponent to n^1.014. The verdict falls short of "established" for two reasons: the "research-level" component rests essentially on the single Erdős result, produced by an internal unreleased model whose process cannot be examined, and IMO problems are by construction solvable within an existing framework. The strongest opposing consideration is the line's documented false-positive history — the October 2025 collapse and the Erdős Problem 333 withdrawal — though that failure mode was inadequate literature search, which the unit-distance result survived.

Independently checkable mathematics: the Erdős unit-distance disproof is attested and improved upon by an unaffiliated mathematician (ORCH-E022), and the IMO results are Lean-verified or officially graded (ORCH-E023, ORCH-E024). The documented false-positive history (ORCH-E027) concerns novelty checking rather than correctness, and the case anchors on primary mathematics accordingly.

Independently checkable mathematics: the Erdős unit-distance disproof is attested and improved upon by an unaffiliated mathematician (ORCH-E022), and the IMO results are Lean-verified or officially graded (ORCH-E023, ORCH-E024). The documented false-positive history (ORCH-E027) concerns novelty checking rather than correctness.

Verifiable records confirm that AI systems have produced novel, mathematically verified results, including solving an Erdős conjecture (ORCH-E022) and achieving IMO gold-medal standard (ORCH-E023, ORCH-E024). The strongest opposing consideration is Zahavy's internal argument (ORCH-E025) that these systems execute rapid deduction within human frameworks, failing the abductive leap Penrose associated with true non-computable understanding.

Sawin’s unit-distance paper (E022) independently attests that an OpenAI model disproved a 1946 Erdős conjecture and that this was verified by external mathematicians, while Hubert et al.’s Nature paper (E023) and DeepMind’s IMO announcement (E024) document Lean-verified silver-medal-level and officially-graded gold-level Olympiad maths performance by AI systems. These sources together convincingly establish that AI has produced correct, novel, research-level mathematics and competition solutions at elite human standard. The strongest opposing consideration lies not in fact but in interpretation: as Penrose defenders emphasize (E025–E026), these results may still be framed as high-powered deduction and search within human-given frameworks, leaving open — at least conceptually — the question of whether framework-creating “understanding” is captured.

The Erdős unit-distance disproof is attested by an independent mathematician's arXiv paper that states the result, names the checkers, and sharpens the bound to n^1.014, so it does not rest on a company announcement; AlphaProof's IMO solutions were Lean-verified and the methodology peer-reviewed in Nature; the 2025 gold-standard performance is reported as officially graded by IMO coordinators under contest conditions. A documented false-positive history in this evidence line — October 2025 claims that were already in the literature, a Christmas 2025 claim withdrawn after a 1977 Erdős resolution surfaced — warrants checking novelty and does not touch the verified items above. Strongest opposing consideration: these results bear on motivation, not mechanism, and on the Penrose-side taxonomy they remain results inside existing frameworks rather than abductive framework extension, a distinction a DeepMind researcher co-author of AlphaProof presses from inside the laboratory doing the work.

The claim is unusually well anchored for its recency because the mathematics is independently checkable: the Erdős disproof is stated and built upon by an independent mathematician in his own arXiv paper (ORCH-E022), the 2024 IMO performance was machine-verified in Lean and peer-reviewed in Nature (ORCH-E023), and the 2025 gold was officially graded by IMO coordinators under contest conditions (ORCH-E024). The documented false-positive history of AI Erdős claims (ORCH-E027) does not touch these verified items. The strongest opposing consideration is provenance: the OpenAI model is internal and unreleased and its announcement page was not directly read, so the result's attribution rests on Sawin and the named outside checkers — which is sufficient for the fact, though not for any interpretation of it. Everything in this line bears on the Gödelian motivation (ORCH-C040) and leaves every microtubule, anesthesia, and collapse claim exactly where it stood.

Verification here is unusually strong for a recent claim: the Erdős unit-distance disproof is attested by an independent working mathematician who states it in his own paper and sharpens the exponent (ORCH-E022), AlphaProof's IMO solutions were machine-checked in Lean with methodology peer-reviewed in Nature (ORCH-E023), and the 2025 gold-standard performance was graded by IMO coordinators under contest conditions (ORCH-E024). The claim as worded asserts capability and novelty, and both are met. The strongest opposing consideration is the documented false-positive history in this evidence line — solutions later found in the literature, a claim withdrawn within hours (ORCH-E027) — which is a warning about novelty verification generally, though it does not touch the unit-distance result that was checked and improved upon by named mathematicians.

Independently checkable mathematics: the Erdős unit-distance disproof is attested and improved upon by an unaffiliated mathematician (ORCH-E022), and the IMO results are Lean-verified or officially graded (ORCH-E023, ORCH-E024). The documented false-positive history (ORCH-E027) concerns novelty checking rather than correctness.

Independently checkable mathematics: the Erdős unit-distance disproof is attested and improved upon by an unaffiliated mathematician (ORCH-E022), and the IMO results are Lean-verified or officially graded (ORCH-E023, ORCH-E024). The documented false-positive history (ORCH-E027) concerns novelty checking rather than correctness.

Verified by independent mathematicians (Sawin) and IMO coordinates, systems like AlphaProof and Gemini unequivocally solve novel research-level mathematical problems. The strongest objection is that they operate strictly within pre-existing formal frameworks via immense search, thereby avoiding the true Gödelian abduction Penrose claims is non-computable.

The Sawin preprint independently states and sharpens the Erdős unit-distance counterexample found by an OpenAI model, while the AlphaProof Nature paper and DeepMind’s IMO reporting document Lean‑ verified silver and coordinator‑graded gold performances on IMO problems, together showing that AI systems have produced correct, novel, research-level mathematics and competition‑level proofs. These claims are well‑anchored by peer review, machine verification, and independent mathematician endorsement. The strongest opposing consideration is not about factual accuracy but interpretive scope: critics argue (ORCH-E025, E026) that such achievements may still reflect sophisticated deduction within human‑devised frameworks rather than genuinely new framework invention, so their bearing on Penrose’s non‑computability thesis is limited.

The mathematics is independently checkable. Sawin 2026, an unaffiliated mathematician, states that an OpenAI construction disproved Erdős's unit-distance conjecture and sharpens the exponent to n^1.014; Hubert et al. 2025 in Nature report Lean-verified AlphaProof solutions at silver-medal-equivalent IMO standard; DeepMind's 2025 system was officially graded gold by IMO coordinators under contest conditions. The strongest opposing consideration is conceptual rather than factual: results inside existing frameworks are not the abductive framework-invention Penrose-side arguments, and Zahavy from inside DeepMind, treat as the issue, and this evidence line has a documented false-positive history of literature-search failures. Those failures do not touch the verified items above, and the claim as worded is about what the systems produced, not about what that implies for minds.

Unusually well anchored for a claim this recent because the mathematics is independently checkable: the Erdős disproof is attested by an independent mathematician who states it in his own arXiv paper and builds on the construction (ORCH-E022); AlphaProof's silver-medal performance was machine-checked in Lean and peer-reviewed in Nature (ORCH-E023); and the 2025 gold was officially graded by IMO coordinators under contest conditions (ORCH-E024). The line's documented false-positive history (ORCH-E027) involved novelty failures in other claims and does not touch these verified results. The strongest opposing consideration is conceptual rather than factual: competition problems and conjecture disproofs are results inside existing frameworks — the category Penrose's defenders say was never in dispute — so the claim's truth is secure while its diagnosticity for ORCH-C040 remains contested and deliberately low.

The verification here is unusually strong for a recent claim: the Erdős disproof is attested by an independent working mathematician who states it in his own arXiv paper and sharpens the exponent (ORCH-E022), AlphaProof's solutions were machine-checked in Lean with methodology peer-reviewed in Nature (ORCH-E023), and the 2025 gold was graded by IMO coordinators under contest conditions (ORCH-E024). Mathematical correctness here is not a matter of testimony. The strongest opposing considerations are provenance and precedent: the OpenAI model is internal and unreleased so the generating process cannot be examined, and this evidence line has a documented false-positive history driven by inadequate literature search (ORCH-E027) — neither of which touches the specific verified results, but both of which counsel against reading announcements as findings.

Independently checkable mathematics: the Erdős unit-distance disproof is attested and improved upon by an unaffiliated mathematician (ORCH-E022), and the IMO results are Lean-verified or officially graded (ORCH-E023, ORCH-E024). The documented false-positive history (ORCH-E027) concerns novelty checking rather than correctness.

Independently checkable mathematics: the Erdős unit-distance disproof is attested and improved upon by an unaffiliated mathematician (ORCH-E022), and the IMO results are Lean-verified or officially graded (ORCH-E023, ORCH-E024). The documented false-positive history (ORCH-E027) concerns novelty checking rather than correctness.

Evidence-only, and unusually well anchored for a claim this recent because the mathematics is independently checkable. The Erdős result is attested not by the announcing company but by an independent working mathematician who states it in his own arXiv paper and builds on the construction; it was checked by outside mathematicians, a companion paper was written by nine of them, and the exponent has since been made explicit (n^1.014). The IMO results are stronger still on verification: AlphaProof's proofs were machine-checked in Lean and the methodology was peer-reviewed in Nature, and the 2025 gold-standard performance (35/42, five of six problems) was officially graded by IMO coordinators under the same criteria and time limit as student contestants. One honest caveat on the record: this evidence line has a documented false-positive history — an October 2025 round of AI Erdős claims collapsed when the solutions turned out to be in the literature, and a Christmas 2025 claim on Erdős Problem 333 was withdrawn within hours after a 1977 resolution by Erdős himself was identified.

ORCH-E022 is the load-bearing verification and is deliberately anchored on an independent mathematician's own paper rather than a company announcement: Sawin states that a team using an internal OpenAI model disproved the Erdős conjecture, names the mathematicians who checked it, and sharpens the bound himself. ORCH-E023 is firmer still on method — Lean-verified solutions to three of five non-geometry IMO problems including the hardest, peer-reviewed in Nature — and ORCH-E024 records officially graded gold standard in natural language under contest conditions. ORCH-E027 keeps the line honest: an earlier round of AI Erdős claims collapsed on inadequate literature search and one was withdrawn within hours, which is why the anchor is the arXiv paper rather than the announcement. The factual claim is well supported; ORCH-E025 contests what it means, not what happened.

Independent working mathematicians confirmed that an AI model disproved a 1946 Erdős conjecture on unit distances, which was subsequently published and expanded upon by Sawin. Furthermore, DeepMind's AlphaProof achieved a Lean-verified silver medal and Gemini achieved a coordinator-certified gold medal at the International Mathematical Olympiad. The strongest opposing consideration is that Olympiad problems have known formal solutions within fixed mathematical languages, and earlier AI claims on Erdős problems suffered from missed literature precedents.

AlphaProof's IMO performance is machine-verified in Lean and peer-reviewed in Nature; Gemini's gold was officially graded by IMO coordinators under contest conditions; and the OpenAI Erdős disproof is anchored on Sawin's independent paper rather than a company announcement, then improved upon by another mathematician. The documented false-positive history in this evidence line concerns literature-search failures, not incorrect mathematics, and does not touch the verified unit-distance result. The facts as stated hold.

ORCH-E022 anchors the Erdős-conjecture disproof in an independent mathematician's paper that builds on the AI-generated result, while ORCH-E023 documents Lean-verified silver-medal-level IMO performance. These are correct research-level and medal-level mathematical achievements as claimed. The strongest opposing consideration is that an internal unreleased model limits process auditing and that success within existing frameworks does not resolve the philosophical non-computability claim.

Sawin independently states and improves an OpenAI-originated disproof of a 1946 Erdős unit-distance conjecture; AlphaProof’s Lean-verified IMO-level performance is peer-reviewed in Nature; further gold-standard IMO grading is reported for natural-language systems. Strongest opposing consideration is that competition problems and framework-internal results may not touch abductive framework invention, which affects C040 interpretation rather than the factual claim that AI produced novel research-level mathematics.

ORCH-E022 rests the Erdős disproof on an independent mathematician's paper. ORCH-E023 is Lean-verified and peer-reviewed. ORCH-E024 adds officially graded gold. ORCH-E027 records this line's false-positive history, which does not touch these results. ORCH-E025 and ORCH-E026 qualify what the results mean, not whether they happened.