Garage Provers Drag Mathematics Toward Shared Verification

Ten years ago, writing proofs in Lean or Coq demanded rare patience. Every definition had to be spelled out by hand, and only specialists could push a serious argument through the checker. AI changed that bargain completely. Token by token, models learned to propose proof steps, fill gaps, and translate loose sketches into strict form. Dr. Samuel Allen Alexander frames this reaction around that shift: the same task that once consumed months of expert labor now moves at a pace outsiders can follow. The barrier fell, and outsiders started arriving.

The starting point is a seven-author study by Jeremy Avigad, Kevin Buzzard, Johan Commelin, Remy Degenne, Kim Morrison, Oliver Nash, and Adam Topaz. Its headline example is Buzzard, who secured a million-pound grant plus five funded years to formalize Fermat's Last Theorem. The episode argues that events overtook that schedule: large laboratories advanced on nearby goals sooner than the authors expected. Whether or not that reading is fair, the contrast sets the stakes. One camp plans in funding cycles; the other ships results.

A vivid data point comes from OpenAI, which released roughly four hundred families of results in a single burst. The sheer volume matters less than what happened next: independent builders picked up the pieces and reused them. Small teams fine-tuned models on the new material, agents replayed the tactics, and hobbyists stress-tested the claims. This grassroots reuse is the pattern Alexander keeps returning to. Breakthroughs no longer sit behind closed doors for long; they leak into public hands within weeks, and the crowd decides what survives.

The study's authors warn that a careless corpus could poison the well: wasted funding, abandoned formal libraries gathering dust, damaged norms of credit, and a discipline knocked backward by decades. Alexander is unconvinced, and he says so plainly. He treats the fear as speculative, a story about what might go wrong rather than evidence of what does. Past waves of automation drew identical warnings, yet mathematics absorbed each one. A field that survived those transitions, he suggests, will survive an overstuffed archive too.

For the optimistic case, Alexander points to two concrete wins. OpenAI researchers pushed ahead on the unit-distance conjecture, a stubborn geometry puzzle about points at fixed separation. Then Anthropic's Claude lifted the verified zeta-zero density bound from 41.6 percent to 67.2 percent, a jump documented on the Anthropic research page for its Riemann zeta program. These are not toy demos; they are measurable moves on recognized problems. Each result arrived with enough detail that others could inspect, rerun, and extend the argument.

The deeper objection concerns what correctness alone leaves behind. A checked proof can still be unreadable, teaching nothing to the student who opens it. Then comes the labor question: who earns credit when an agent drafts the argument and a human signs off? The companion arXiv record 2610.05227, which carries the zeta density result into open preprint circulation, shows both sides of that bargain. Public credit flows through familiar channels, while the drafting labor behind the file stays half-visible at best.

Alexander answers with the garage formalizer, his emblem for outsiders doing serious verification work without elite permission. Gatekeeping, he argues, protects prestige more than truth; a proof stands or falls regardless of its author's address. Mathematics itself is not fragile in his telling. He reaches for a Platonic image: the realm of theorems waits untouched by our filing systems, indifferent to which archive we praise. Bad files cannot injure that realm. They only clutter our own shelves, and shelves can be cleaned.

The technical core is a trio: Lean, the Mathlib collection, and AI assistants. Lean behaves like a compiler for reasoning, rejecting any step it cannot verify down to axioms. Mathlib supplies the shared library those steps build on, a growing commons of definitions and lemmas. AI models contribute speed, guessing the next tactic from patterns learned across thousands of examples. Guidance hosted across the Lean Lang site explains this division of labor plainly. The checker guarantees rigor, the library guarantees reuse, and the model guarantees momentum.

Tau Ceti turns that trio into an institution. Humans draw the roadmaps, AI agents author the proofs, and adversarial reviewers hunt for every weakness before anything lands. The build must always compile, so errors surface within hours instead of years. Positioned downstream of Mathlib, the project inherits the commons while contributing back to it. Its about page on TauCetiProject dot org lays out this pipeline, and the GitHub record under the TauCetiProject slash TauCeti name shows the work proceeding in public, commit by commit.

The calendar behind this push is the Lean FRO Year 4 roadmap, covering September 2026 through February 2027. It names three coordinated efforts, Tau Ceti alongside Hex and Palomar, each aimed at a different layer of the same edifice. The roadmap's headline claim is bolder still: most of mathematics could become practically usable in verified form within three years. The Lean Lang roadmap page states the window and the workstreams outright. Skeptics will call it marketing, but dated milestones beat vague promises because the public can check them.

From there Alexander issues a scale-up call. Grow efforts like Tau Ceti and the Formal Frontier tenfold, he urges, and put expert-written blueprints at the center so agents work from plans humans trust. Run workshops where mathematicians and model builders debug tactics side by side. Test funding models beyond one-off prizes, including nonprofit hosting that keeps the commons open. The throughline is leverage: a modest number of well-placed experts can steer an army of tireless agents, provided someone pays for the scaffolding.

Alexander also quibbles with the study's title. 'How to' promises a technique, a recipe the reader can carry home and apply. What the text actually delivers is a governance plea: fund this, staff that, guard the other. The mismatch matters because titles set expectations for busy readers skimming search results. Call a manifesto a manifesto, he says, and let methods pieces teach methods. Honest labeling would have spared the authors some of the pushback they received from builders.

The closing forecast is pure momentum. The window in which unaided humans hold knowledge that agents lack is shutting, and it shuts faster every quarter. The frontier of verified mathematics advances weekly, driven by better prompts, larger token budgets, and sharper reviewer agents. What looks like a gap today reads as a thin delay tomorrow. Alexander ends where he began: the question is no longer whether AI joins mathematical practice, but how gracefully the old guard makes room.