01The short version
A machine system produced a large body of mathematics at research-level ambition. Some of it may hold up. Most of it is unverified. The formal checking done so far covers a minority of the corpus. The real signal is about how mathematics gets made, not about any single theorem.
02What dropped
722 manuscripts. 372 result families. One initial commit. A catalog file lists 235 formalizations against those 722 manuscripts. Only 10 of 372 families ship with abridged reasoning traces. The repo keeps old versions accessible and records corrections as new versions. That versioning choice matters. It makes the corpus auditable.
Lean here means a proof assistant. It is software that checks each logical step of a proof. A formalization means the result has been translated into a form Lean can check. 235 of 722 manuscripts have that. The rest stand on human-readable PDFs alone. The repo itself warns that unformalized results could have issues.
03The ranked notables
These are the claims our team ranked highest, in rough order of weight.
A near-Riemann result, in two versions. The Riemann hypothesis is the famous open claim that the prime numbers follow a precise hidden pattern. The September 30 manuscript claims every Dirichlet L-function has no zeros past 7/8. A Dirichlet L-function is a number-theoretic function in the family that includes the Riemann zeta function. That means primes are nearly as regular as the full hypothesis would imply. A companion October 5 manuscript proves a weaker 11/12 bound with human editing help. The stronger 7/8 paper has Lean coverage. The 11/12 companion does not. It reads as a checkable alternate route, not a replacement.
Hodge for CM abelian varieties. The Hodge conjecture is a Clay Millennium problem. It asks whether certain geometric patterns (Hodge classes) always come from algebraic shapes. The manuscript claims a full proof for one large class (CM abelian varieties, highly symmetric geometric objects), plus consequences for related conjectures. It has no Lean coverage. It is one of the repo's stated exceptions to the standard machine procedure. Several sibling papers cite it as a prerequisite, so a fault here would spread.
Unique Games proved. The Unique Games Conjecture was the central open problem in hardness of approximation. That field studies which optimization problems can be solved well and which cannot. The manuscript claims a full proof, plus direct hardness results for problems like Max-Cut and Vertex Cover. It has Lean coverage. Of the eight ranked notables (the speed pair below is technique watch), this one has the most reusable machinery for optimization work.
Free group factors isomorphic. Von Neumann algebras are structures from functional analysis used in quantum theory. The manuscript claims two famous examples built from free groups are the same object, which would settle a decades-old question. Lean documentation exists for part of the claim. The catalog entry is missing, which needs cleanup.
Hilbert's tenth problem over Q. Hilbert's tenth problem asks whether an algorithm can decide if a polynomial equation has a solution. The manuscript claims no such algorithm exists for rational solutions. That would settle a major open problem in logic. It has no Lean coverage.
Catalan's constant is irrational. Catalan's constant is a famous number (about 0.916) whose nature was unknown. The manuscript claims it cannot be written as a fraction. Lean documentation exists for the core claim. The catalog entry is missing here too.
Both Mahler conjectures. Convex geometry studies the shapes of convex bodies. The Mahler conjectures give the smallest possible volume of a shape paired with its dual. The manuscripts claim both the symmetric and general cases, plus equality cases and a symplectic width result. This is the best-evidenced item. It has Lean statements, solution files with no placeholders, ten exact-arithmetic check programs, and a reasoning trace. The trace itself disclaims end-to-end linkage between its own parts. The bridge between the Santalo-point and infimum formulations rests on one unverified product-infimum step.
EVIDENCE: STRONGEST IN CORPUS — LEAN STATEMENTS + CHECK PROGRAMS + TRACEShortest superstring at factor 2. Given a set of strings, the shortest common superstring is the shortest string containing every input string. It models genome assembly. The manuscript claims a deterministic algorithm, running in polynomial time, whose output is never more than twice the optimum. Deterministic means no random choices. Polynomial time means the running time grows manageably with input size.
Human work took the best ratio from 3 in 1991 to 2.466 in 2023 to 7/3 in September 2026. The model reached 2, the bound conjectured for the greedy algorithm since 1988, without using greedy at all. A Lean formalization scope document exists with a comparator statement. The human 7/3 paper and the model's preprint are both dated September 24, 2026.
Faster multiplication, plus a Fourier companion. One manuscript claims integer multiplication below n log n on a fixed multitape machine. That is a model of computation with fixed tapes instead of random access memory. The claimed saving is tiny (an exponent shaved by 2 to the power minus 182) with enormous constants. A companion manuscript claims a similar saving for the exact discrete Fourier transform in a complex-register model. The Fourier transform is the standard tool for converting signals into frequencies. Neither manuscript is formalized. A related finite-network paper is formalized, but that covers only the core gadget, not the full proofs.
FORMALIZATION: NONE ON EITHER FULL PROOF04What our investigation actually found
The formalization map is partial. 235 catalogued formalizations against 722 manuscripts. Lean-covered families are candidates to build on. Unformalized ones are candidates to read and check. That split held across every notable we examined.
Mahler is provisional, not cleared. The four Lean statements match the paper theorems. The entry-point proof files contain real proofs with no placeholders. The ten check programs read as sound exact-arithmetic certificates. But three gaps block a clean verdict. The runner needs manifest and baseline files that are absent at this commit, so the computation cannot be rerun as shipped. The specification file the programs cite is absent. And the general-case formalization exists on disk but is missing from the catalog index, so catalog-driven tooling would skip it. The paper PDFs themselves still await line-level review.
Both speed claims share one linchpin. The multiplication and Fourier manuscripts rest on the same fixed-size combinatorial construction with parameter h equal to 100. One rank-counting argument produces the saving in both papers. If that count fails, both exponents fail with it. Check it once and both results move together.
The survival call favors H10 over Hodge. Of the two unformalized flagships, the logic result is likelier to survive than the geometry result, in our judgment. The logic proof concentrates its novelty in one inequality plus standard machinery. The geometry proof chains five interdependent novelties, each of which must hold at full strength, inside a single manuscript. The geometry failure would also pull down four dependent papers. Both remain read-and-check.
05What it means downstream
| Area | Our judgment |
|---|---|
| Theory | Genuine movement if the proofs hold. Counterexamples and hardness results transfer first. |
| Practice | No change. Constants are galactic. No library, chip, or pricing formula is affected. |
| Finance | Method transfer only, in our judgment. Optimization and sampling machinery may speed up prototyping. Nothing here touches a production model. |
| The real signal | Machine systems can now propose full research programs at scale. Verification bandwidth is the binding constraint. |
On finance, to be plain: this corpus says nothing about markets. Any finance effect is our inference from the capability signal, not a source claim. The plausible shift is labor toward specifying problems and checking machine derivations. Model-risk teams could copy the Lean-catalog pattern for pricing lemmas. Proprietary math edges get easier to reproduce, so durable edge moves to data and verified infrastructure. None of that changes a price today.
06What we still do not know
- Correctness of any single manuscript. That needs expert review and finished formalizations.
- Selection bias. About 4,000 problems were posed and 722 manuscripts published. The filter is undisclosed.
- Coverage trajectory. The repo promises more formalizations with no schedule.
- Trace fidelity. Ten abridged traces across 372 families cannot show how the search really ran.
- Community reception. No referee reports exist yet. The version history is empty on day one.
07The acceleration thesis
The review above covers what was published. This section is about what changes if even a fraction of it holds, and what has already changed regardless.
Verification is now the scarce resource. For centuries, finding a proof was the hard part and checking it was routine. That ratio inverted on October 6. Refereeing has to industrialize: Lean coverage first, targeted human review second, prestige last.
Falsifiable claims win. A 70-page proof is hard to check. A single finite inequality that the whole proof funnels through is easy to aim at. Future machine-generated mathematics will be judged partly by whether its load-bearing claim is small.
"Optimal" gets repriced. A 55-year-old speed limit fell to a short compute run. Every "we believe this is the best possible" in computer science now carries a discount. The cost of searching for the better idea just collapsed.
Results compound. Seven hundred proofs landing at once means methods migrate. Progress becomes combinatorial across results rather than additive. The corpus is a shared parts bin. The most valuable outputs may be the techniques, not the theorems.
The loop compresses. Conjecture to candidate to machine-checked verdict used to take decades. It now takes days for the candidate and an unknown but shrinking time for the verdict. The question stops being "can we solve it" and becomes "how fast can we check the solution."
The counterweight: none of the 722 is confirmed. The acceleration so far is entirely in candidate generation. One lab, one unreleased model, one undisclosed filter over roughly 4,000 posed problems. Being early cuts both ways.
What would confirm the thesis. A second lab reproduces the pipeline. Lean coverage grows on a visible schedule. One falsifiable core gets killed or confirmed in public. If those happen, the shifts above stop being predictions.
08Sources
Repo, commit, license, and counts are pinned in the opening paragraph. Family and catalog pointers (CONTENTS, formalization catalog, family docs, preprint directories) are all read at that commit. Finance effects in this post are judgment. Manuscript claims are unverified day-one material.