{"schema":"sanctity-signal/0.2","record_type":"signal","id":"google-antigravity-teamwork-proofs-e094","domain":"science","drop_id":"signal-e094","occurred_at":"2026-11-13","how_to_use":"This is a structured record of one real-world action and its competing interpretations, from Sanctity Signal. To build content from it (a reel script, a blog, a thread): open on the hook, ground the viewer with plain and observed, walk the dated trail, present the interpretations as live tension between named readings, and close on the before/after shift and what_to_watch. Hard rules: items under established are fact; items under unknown are never presented as fact; the evidence status (verified or reported) is stated, not dropped; sources are cited.","actor":"Google Research","action_type":"publication of machine-checked formalizations across 6 open mathematical problems, with an offer of 6 months of top-tier access to 10,000 academic researchers","observed":"Vahab Mirrokni, Google Fellow and Vice President at Google Research, published \"Advancing science and engineering with Antigravity's /teamwork\" on research.google on November 13, 2026. The post reports results from /teamwork, a multi-agent feature in Antigravity, run with human experts across mathematics, chemistry, systems software and machine learning engineering. On the Courtade-Kumar conjecture, an information theory problem posed in 2013, it says Gemini 4 Argon and earlier Gemini models developed most of the new analytical bounds, and that \"the team formalized the entire proof in Lean 4, producing 50 million lines of code.\" On the Weil conjectures it says /teamwork-LEAP fully autonomously produced an end-to-end machine-checked formalization: \"This formalization spans about 1.8 million lines of Lean, produced across 28,800 subagent invocations.\" On Hilbert's tenth problem, \"the system also synthesized over 919,000 lines of Lean to produce the first formalization\" of a recent proof over general rings of integers. It reports the first formalization of the Birkhoff conjecture, a dynamical systems problem open since 1927, and of the 3D Bochner-Riesz conjecture, open more than 50 years, \"finding and fixing a gap in a recently released informal proof along the way.\" It reports an independent reproduction of an end-to-end formalization of a recent proof of finite-time blowup for 3D Navier-Stokes with smooth forcing, which it describes as open for about 90 years. Further results are named on Tarjan's Deque Conjecture, Boolean matrix multiplication and integer multiplication. On peer review, \"authors of 578 submissions (over a third of all submissions)\" to STOC 2027 opted in to Deep Review; in an ongoing survey of 70 respondents, 90% rated the feedback mostly or almost all factual and 90% said they would use it again. Gemini 4 Argon is named as the main model, with Carbon and Hydrogen in the introduction. On access: /teamwork, /boost and Deep Review are available to external users now, Deep Review gives 2 free calls per user each week, and the team is \"giving 10,000 academic researchers six months of free access to our highest usage tier\" by application. Not established: compute cost, wall-clock time, how many of the 10,000 places get filled, eligibility, and whether anyone outside has rerun the work.","corrections":[],"evidence":{"status":"verified","note":"Verified: the post was read directly at research.google, and every number and quotation here comes from it. That establishes publication, not reproduction. No outside group has rerun the formalizations, and the Lean files were not located. The dated event that would move it: an independent recheck by June 30, 2027.","sources":[{"role":"primary","label":"Google Research: Advancing science and engineering with Antigravity's /teamwork, by Vahab Mirrokni, Nov 13, 2026","url":"https://research.google/blog/advancing-science-engineering-with-antigravitys-teamwork-feature/"},{"role":"corrob","label":"Interesting Engineering: Google's Antigravity agents formalise major mathematical proofs","url":"https://interestingengineering.com/ai-robotics/google-antigravity-multi-agent-research"},{"role":"corrob","label":"Datastudios: Antigravity /teamwork, 6 months free top-tier access for 10,000 researchers","url":"https://www.datastudios.org/post/google-antigravity-teamwork-free-top-tier-access-10000-researchers-six-months"},{"role":"corrob","label":"FourWeekMBA: Google Gemini agents formalize 99-year-old Birkhoff problem","url":"https://fourweekmba.com/ai-google-gemini-agents-formalize-99-year-old-birkhoff-problem/"}]},"strength":{"cost":"High","commitment":"High","reversibility":"Moderate","persistence":"Years","rationale":"Cost is high. The post's own numbers imply very large compute: 28,800 subagent invocations for one formalization, 50 million lines of Lean for another, and 10,000 researchers on the highest usage tier for 6 months, given away rather than sold. Commitment is high: a named Google Fellow and Vice President published the claims under his own byline, with the access offer live. Reversibility is moderate. The seats run out in 6 months and the programme can quietly end, but published formalizations stay checkable forever. Persistence is years: formal proofs do not decay, and the review experiment has already drawn a third of a conference's authors."},"shift":{"prior":"AI helps mathematicians draft and check parts of their proofs.","action":"Vahab Mirrokni, Google Fellow and Vice President at Google Research, published \"Advancing science and engineering with Antigravity's /teamwork\" on research.google on November 13, 2026. The post reports results from /teamwork, a multi-agent feature in Antigravity, run with human experts across mathematics, chemistry, systems software and machine learning engineering. On the Courtade-Kumar conjecture, an information theory problem posed in 2013, it says Gemini 4 Argon and earlier Gemini models developed most of the new analytical bounds, and that \"the team formalized the entire proof in Lean 4, producing 50 million lines of code.\" On the Weil conjectures it says /teamwork-LEAP fully autonomously produced an end-to-end machine-checked formalization: \"This formalization spans about 1.8 million lines of Lean, produced across 28,800 subagent invocations.\" On Hilbert's tenth problem, \"the system also synthesized over 919,000 lines of Lean to produce the first formalization\" of a recent proof over general rings of integers. It reports the first formalization of the Birkhoff conjecture, a dynamical systems problem open since 1927, and of the 3D Bochner-Riesz conjecture, open more than 50 years, \"finding and fixing a gap in a recently released informal proof along the way.\" It reports an independent reproduction of an end-to-end formalization of a recent proof of finite-time blowup for 3D Navier-Stokes with smooth forcing, which it describes as open for about 90 years. Further results are named on Tarjan's Deque Conjecture, Boolean matrix multiplication and integer multiplication. On peer review, \"authors of 578 submissions (over a third of all submissions)\" to STOC 2027 opted in to Deep Review; in an ongoing survey of 70 respondents, 90% rated the feedback mostly or almost all factual and 90% said they would use it again. Gemini 4 Argon is named as the main model, with Carbon and Hydrogen in the introduction. On access: /teamwork, /boost and Deep Review are available to external users now, Deep Review gives 2 free calls per user each week, and the team is \"giving 10,000 academic researchers six months of free access to our highest usage tier\" by application. Not established: compute cost, wall-clock time, how many of the 10,000 places get filled, eligibility, and whether anyone outside has rerun the work.","update":"Agent teams run whole formalizations end to end, and find human errors."},"inference":{"established":["The post claims 1.8 million Lean lines for the Weil conjectures, autonomously produced.","Authors of 578 STOC 2027 submissions, over a third, opted into Deep Review.","Google is giving 10,000 academic researchers 6 months of top-tier access."],"unknown":["Independent verification. No outside group has rerun these formalizations yet.","New mathematics. Most results formalize proofs humans had already written.","Cost. The post gives no compute figure for any of the runs."]},"narrative":{"hook":"Google's agents machine-checked proofs of 6 hard problems, then opened the tool to 10,000 researchers.","plain":"Google Research published results from Antigravity's /teamwork agents on November 13, including 1.8 million lines of Lean for the Weil conjectures. Caveat: most results formalize existing proofs rather than prove new theorems.","story":["Formal verification was the bottleneck in mathematics: one hard proof took a human team years to machine-check. Google says one system produced 1.8 million lines of Lean with no human in the loop, and fixed a gap in a recent informal proof. The 10,000 free seats turn a demonstration into a dependency."],"trail":[{"d":"1927 TO 2013","t":"The Birkhoff conjecture is posed in 1927, Bochner-Riesz in the 1970s, Courtade-Kumar in 2013. All stay open. Machine-checking a single hard proof takes a human team years."},{"d":"NOV 11","t":"Google releases a DNA model that reads genes cell by cell, free for research. Edition 087's science card."},{"d":"NOV 12","t":"Google releases Gemini 4 Hydrogen for coding and agent workflows. Edition 090's card. Antigravity is where those agents run."},{"d":"NOV 13","t":"Google Research publishes the /teamwork results: 6 named problems, 1.8 million Lean lines on the Weil conjectures alone, a gap found in a recent informal proof, and 10,000 free academic seats."},{"d":"NOV 13","t":"The same day a Helsinki logician tells the Guardian that Claude turned 1,009 pages of Godel's shorthand into 663 theorem candidates, about 400 of which checked out. Two laboratories, one week, formal mathematics."},{"d":"Reacts","r":true,"t":"Lean's maintainers, whose checker is now the gate on agent output; the 578 STOC 2027 authors who opted in to machine review; journal and conference referees; Anthropic, whose model did the Godel work the same week; and the 10,000 academics who will apply."}],"glossary":[{"term":"Lean 4","def":"A proof checker. You write a proof in its language and the software either accepts every step or rejects it, which is why a machine-checked proof cannot be bluffed."},{"term":"Formalization","def":"Rewriting a human proof in a checker's language. It usually takes far longer than the original proof, which is why most published mathematics is never formalized."},{"term":"Subagent invocation","def":"One call to a model acting as a worker inside a larger agent run. 28,800 of them means the job was split into that many small pieces."},{"term":"Weil conjectures","def":"Four statements tying the counting of equation solutions to geometry, posed in 1949 and proved by 1974. The proofs are long and hard, which makes them a stress test."}]},"judgment":{"prompt":"Which explanation do you believe?","interpretations":[{"id":"a","label":"Formal proof just got automated","text":"6 formalizations in one post, including one with no human in the loop.","abstain":false},{"id":"b","label":"A showcase for Gemini 4","text":"Google needs a reason to put Gemini 4 in front of working researchers.","abstain":false},{"id":"c","label":"Lean was the real unlock","text":"The checker makes agent output safe to trust. The models are incidental.","abstain":false},{"id":"d","label":"Normal progress in formalization","text":"Machine-checked proofs have grown steadily for years. This is the next step.","abstain":false},{"id":"e","label":"Too early to call.","text":"The evidence cannot yet separate these.","abstain":true}],"distribution":{"status":"held back until you have made your call","calls":null}},"story_thread":null,"what_to_watch":[{"linked_interpretation":"a","direction":"strengthens","observable_event":"An outside group rechecks the Weil or Hilbert formalization and confirms it, or a second laboratory reports an autonomous end-to-end formalization of a comparable open problem, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"},{"linked_interpretation":"a","direction":"weakens","observable_event":"A substantive error is found in one of the 6 named formalizations, or Google withdraws or materially corrects the post, by March 31, 2027.","window_end":"2027-03-31","state":"unresolved"},{"linked_interpretation":"c","direction":"strengthens","observable_event":"A journal or a major conference adopts machine-checked submission or machine-assisted review as stated policy, by December 31, 2027.","window_end":"2027-12-31","state":"unresolved"}]}