By Paul Derrington, Jr.
CEO | Chief Engineer | Research Engineer | Editor, D-CSIL
5 min read
Key Takeaways
Google Research's November 13 /teamwork report says teams using its agents formalised a 2013 conjecture in 50 million lines of Lean. A November 14 count by the AI news site Sysnote matched that number, and the repository's own guide says under 0.5% of it is readable proof code; the rest is generated certificate data checked by Lean's kernel.
50,373,106 lines — counted, not quoted
Google Research published its /teamwork report on November 13. The headline number was 50 million lines of Lean 4 for a formalised proof. The AI news site Sysnote said on November 14 that it cloned the public repository — dpwoodru/general-courtade-kumar-lean — and counted 45,510 .lean files and 50,373,106 lines. That matches the repository's own source manifest (45,500 modules, 50,371,955 lines) plus ten auxiliary Lean files, so Google's rounded number holds.
The repository's own guide says only 1,359 files — 238,690 lines — are readable proof code. The other 44,141 files, 4.88 GB, are generated certificate data: numerical bounds over pieces of the parameter space that the Lean kernel checks, mostly with decide. By line count, under half a percent is readable proof code; the guide says the certificate files are written to be checked, not read.
The proof underneath was from September
The conjecture, posed by Thomas Courtade and P. R. Kumar in 2013, asks a plain question: when a random bit string passes through a noisy channel, can compressing it to a single bit preserve more information than a dictator function, which just copies one input bit? The team's proof, formalised in Lean, says the answer is no. Researchers from Google, CUHK, USC and Carnegie Mellon used /teamwork's Long Proof pattern. The CUHK team proposed a candidate function that reduced the problem to a 14-variable optimisation, and Google says Gemini 4 Argon and earlier Gemini models partitioned the parameter space, developed most of the new analytical bounds and extended the proof from balanced to unbalanced functions without human input.
The proof itself is not new in November: its preprint went up on September 21, and Google's post links to it. Sysnote notes other proofs of the same conjecture appeared the same day and in the weeks after — none of them mentioned in Google's November post. Google's version comes with a formal audit trail. The repository says a human needs to check only the 63-line statement file and the Mathlib definitions it uses, and Google's post notes that a Lean proof settles a problem only once human mathematicians confirm the formal statement matches it. The repository's README records a clean rebuild on September 23 with a byte-identical final artifact, acceptance by Lean's comparator against the Formal Conjectures statement on October 4, and acceptance of all 22,107,215 declarations by con-leche, the Lean FRO's external checker, on October 10. It says the final theorem uses only Lean's three standard axioms, with no sorry and no added axiom.
1.8 million lines with no public code yet
Google also says /teamwork-LEAP, paired with Gemini 4 Argon, fully autonomously produced an end-to-end, machine-checked formalisation of the Weil conjectures — the landmark results on counting solutions of polynomial equations over finite fields, proved by Pierre Deligne in 1974. Google puts that formalisation at about 1.8 million lines of Lean, produced across 28,800 subagent invocations.
That claim cannot be checked yet. Google's post links to Wikipedia pages for the Weil conjectures and Deligne, with no repository, no paper and no statement file, and Sysnote found no Weil or LEAP repository under Google's GitHub organisations. Until the code is out, Sysnote notes, outsiders can't see which theorem it states, which axioms it uses, or whether it rebuilds. A week earlier, on November 6, Axiom published its own Weil formalisation publicly under Apache 2.0; Sysnote counted 213,159 lines in that repository.
The same pattern, pointed at molecules and code
In chemistry, Google says it organised 30 specialised agents into a self-correcting loop with a physics-informed spectral simulator to determine molecular structures from nuclear magnetic resonance data. It reports 95% top-1 exact match and 100% top-3 accuracy on complex benchmark molecules, nearly doubling single-agent performance. Those are Google's reported results on its benchmark.
In code, Google says /teamwork rewrote libgav1, its AV1 video decoder, from C++ into around 170,000 lines of Rust, with 100% bit-exact parity across a broad conformance test suite. The first port ran about four times slower than the original; after the agents were given profiling tools, Google says the Rust version runs 20–30% faster than the original C++ and hand-optimised assembly on both x86_64 and ARM. A separate two-week campaign rewrote the remaining C++ in the Zircon microkernel, adding about 800,000 lines of safe Rust that Google says pass all kernel unit tests and boot on arm64. Sysnote notes no public code is linked for either Rust port, so both rest on Google's account for now.
What is available now
The design behind the results is consistent: /teamwork pairs parallel exploration with dedicated falsifiers — the Lean 4 checker, unmodified test suites, execution simulators, web search or hardware counters — and gives each specialised agent only the files and tools its step needs, while critics and auditors check results from a clean context.
Google says external scientists and developers can use Gemini 4 models with /teamwork and /boost in Antigravity, and Deep Review, today, and that Deep Review users get two weekly calls free. The formal proof pattern is not in that list: Google says upcoming updates will bring it and other specialised workflows into Antigravity. Applications are also open for a programme giving 10,000 academic researchers six months of free access to Google's highest usage tier.