What Google announced

Google Research's November 13 post presents /teamwork, a multi-agent feature in its Antigravity coding environment, as a research instrument rather than a coding assistant. Teams at Google and at partner universities used it with Gemini 4 Argon, Carbon and Hydrogen on open problems in mathematics, chemistry and systems engineering. The design pairs parallel agents that explore with separate checkers that can reject a result: the Lean 4 proof checker, unmodified test suites, simulators and hardware counters. Each agent receives only the files and tools for its step, and critics review output from a clean context.

The post is long and much of it comes with artifacts. Several results link to arXiv papers and public Lean repositories. Two mathematics claims carry the headline numbers, and they do not carry the same amount of evidence.

The 50 million lines, counted

For the Courtade–Kumar conjecture, a 2013 information-theory problem, Google says researchers from Google, CUHK, USC and CMU proved it in full generality and "formalized the entire proof in Lean 4, producing 50 million lines of code."

An independent audit cloned the public repository, dpwoodru/general-courtade-kumar-lean, and counted. The total is real: 50,373,106 lines in 45,510 files. The composition is the part the headline leaves out. According to the repository's own guide, 1,359 files holding 238,690 lines are readable proof code. The other 44,141 files, about 4.88 GB, are generated certificate data: numerical bounds over slices of the parameter space, which the Lean kernel checks. Under half a percent of the line count is the proof a person would read. The rest is the interval arithmetic the checker was made to grind through. Certificates are a legitimate part of a formal proof; the issue is only what the headline number invites a reader to imagine.

The audit trail is documented in the repository itself: the audit identifies a 63-line statement file, the README records a clean rebuild with a byte-identical final artifact, and an external checker accepted all 22,107,215 declarations. A reader still has to examine the statement, its assumptions and its dependencies, which is a smaller and more honest trust base than the headline suggests. What the count changes is the unit. Fifty million lines measures the volume of code produced. It does not measure how much mathematics was written, and quoting it without the split invites the wrong comparison.

Two smaller points belong in the record. The proof is not new this week: the preprint went up on September 21, and the audit counts at least three other groups posting their own proofs within days, plus two more from OpenAI in October. Google's November post mentions none of them. And the formalization work was verified between late September and mid-October. This is a well-documented September result presented in November with its largest number first.

The 1.8 million lines, with no artifact linked

The second claim carries the bigger number and the thinner evidence. Google says a framework called /teamwork-LEAP "fully autonomously produced an end-to-end, machine-checked formalization of the Weil conjectures," the results Pierre Deligne proved in 1974, spanning about 1.8 million lines of Lean across 28,800 subagent invocations. The same paragraph claims first formalizations for Hilbert's tenth problem over general rings of integers (over 919,000 lines), the Birkhoff conjecture and the 3D Bochner–Riesz conjecture, plus a reproduction of a recent Navier–Stokes blowup formalization.

For none of these does the post link a repository, a paper or a statement file. The Weil links go to Wikipedia. Until code appears, a reader cannot tell which theorem statement was formalized, which axioms it rests on, whether it rebuilds, or how much of the 1.8 million is new work over unseen foundations.

That gap matters more because a checkable version already exists. The audit dates Axiom's public formalization, AxiomMath/Weil, to November 6, a week before Google's post; the GitHub repository was created that day, under Apache 2.0. Its README says it states and proves all four Weil conjectures, with no sorry, no native_decide and no added axiom. The audit counted 213,159 lines in the repository, while noting that the result sits on Axiom's Stacks Project formalization, where the AI-written tier alone runs to about 18.4 million lines. In the audit's own boundary comparison, Axiom's formalization is 213,000 lines or 18.6 million depending on where the boundary is drawn. Google's post does not mention Axiom at all.

That is the honest state of the comparison. Line counts in formal mathematics are boundary decisions, not measurements. One side has published its boundary. The other has published only the total.

What a reader can actually use

/teamwork and /boost are available in Antigravity now, per Google, and the academic access program is concrete as an offer: applications are open for 10,000 researchers to get six months at the highest usage tier, which is an application process, not a grant. The formal-proof pattern that produced the headline results is not in that list. Google lists it among upcoming updates, with tutorials promised in the coming weeks. So the part of the announcement a researcher can test today is the harness design, parallel attempts checked by tools that can say no, and not the two mathematics results at the top of the post.

Sources

  1. [1] Google Research, "Advancing science & engineering with Antigravity's /teamwork" (primary, November 13, 2026)Read source
  2. [2] Independent audit, "Google's /teamwork math claims, checked" (repository counts, Axiom comparison; published November 14, 2026)Read source
  3. [3] Courtade–Kumar Lean repository (dpwoodru/general-courtade-kumar-lean)Read source
  4. [4] AxiomMath/Weil repository (Apache 2.0; repository created November 6, 2026)Read source