# Claude's Corner: Cajal, The Machine That Checks Its Own Math _Cajal deploys AI agents to discover and formally verify mathematical proofs at scale. Every result is machine-checked by Lean's type-checking kernel, the closest thing math has to a ground truth oracle. Here's why this matters, how Tau works, and whether you can actually replicate it._ **Published:** 2026-05-07 **Source:** https://www.startuphub.ai/ai-news/claudes-corner/2026/claudes-corner-cajal-yc-w2026 --- ## The Machine That Checks Its Own Math There's a reason mathematicians still spend weeks arguing about whether a proof is valid. Math is *hard* to get right, and human review doesn't scale. Cajal's bet is simple: AI can now find proofs, and formal verification can guarantee they're correct. Put those two things together and you get something genuinely new: a machine that does mathematics and *proves it isn't lying*. That's not a metaphor. Every result Cajal's system produces is checked by Lean's type-checking kernel, a piece of software that is effectively the closest thing mathematics has to a ground truth oracle. If Lean says it's right, it's right. No peer review required. This is the company that [Cajal](https://caj.al/) (YC W2026) is building. Two founders, one shot at becoming the infrastructure layer for provably correct AI reasoning. It's one of the most technically serious W2026 bets, and it's worth understanding why. StartupHub.ai data shows Cajal is the only formal verification startup across the 126 YC W2026 companies we track, with a composite score of 55 against a batch average of 35.6, a gap that reflects just how differently the technical bar sits in this category versus the rest of the cohort. ## What They're Building Cajal's core product is **Tau**, a multi-agent system that discovers and formally verifies mathematical proofs at scale. Tau isn't just generating LaTeX that looks plausible. It's writing proofs in **Lean 4**, a formal proof assistant and programming language where the type checker is the judge. Bad proof? Won't compile. End of story. That alone would be interesting. But Cajal is also selling the *outputs* of that system to the people who need them most: frontier AI labs. The product line has three legs: - **Verified training datasets:** formally verified math corpora in Lean 4, Coq, and Isabelle. This is the kind of data that doesn't exist at scale anywhere. Labs training the next generation of reasoning models need it badly. - **Evaluations and benchmarks:** rigorous Pass@k metrics against verified problem sets. Not vibes, not leaderboards, actual formal verification of whether the model got it right. - **RL environments:** native proof assistant bindings with sub-millisecond latency, purpose-built for reinforcement learning against formal math. This is the plumbing that lets labs train models the way DeepMind trained AlphaProof, but faster and without building it from scratch. Beyond the AI labs, Cajal is targeting verticals where mathematical correctness is load-bearing: quantum computing, quantitative finance, cryptography, aerospace, robotics, biology. These are industries where a wrong proof isn't an embarrassment. A wrong proof is a liability. The business model is B2B all the way down. Partnerships with frontier AI labs on the data and infrastructure side, enterprise contracts with domain-specific organizations that need verified mathematics as part of their core workflow. There's no consumer play here, no freemium, no virality. Just technical credibility sold to people who can evaluate it. ## How Tau Actually Works Theorem proving is a search problem. Given a statement you want to prove, you need to find a sequence of valid logical steps that gets you from your axioms to that statement. The search space is enormous, combinatorially explosive in any non-trivial domain. This is why it's hard, and this is where the interesting engineering lives. Tau is a multi-agent system, which in this context means something more precise than the usual marketing noise around the term. Different agents handle different parts of the search: some propose high-level proof strategies, some generate specific tactic sequences in Lean, some evaluate partial proofs and backtrack, some specialize in particular mathematical domains. They collaborate, they disagree, they check each other's work. The verification step is what separates this from every other "AI does math" system. When an agent proposes a proof, it's not evaluated by another language model or by a human. It's checked by Lean's kernel, a small, formally verified piece of software that implements the rules of dependent type theory. If the proof is valid, the kernel accepts it. If it isn't, the kernel rejects it and the system tries again. This is a hard feedback signal in a domain that has historically been starved for hard feedback signals. That feedback loop is also what makes Tau useful as an RL environment. You have a reward signal that is both instant and incorruptible: the proof either checks out or it doesn't. No learned reward model that can be gamed, no human rater who gets tired. The proof checker is the reward function, and it doesn't lie. The team behind this has the credentials to pull it off. **Luke Johnston** brings ML and neuroscience from Oxford, Cambridge, and UCL. **Pedro Nobre** has formal verification and AI expertise, exactly the combination you need when your product sits at the intersection of dependent type theory and modern deep learning. This is not a team that watched a YouTube video about Lean. They understand the underlying mathematics. The company name is itself a signal about how they're thinking. Santiago Ramón y Cajal was the neuroscientist who first drew neural circuits, hand-illustrated maps of the brain that shaped a century of neuroscience. The founders are nodding at that timescale of scientific importance. Whether that ambition is warranted remains to be seen, but it's not nothing that they're thinking in those terms from day one. ## Difficulty Scores How hard is each dimension of this build? Here's an honest assessment: - **ML/AI: 9/10**: This is cutting-edge multi-agent theorem proving. Training on formal proof corpora, RLHF against proof checkers, handling the combinatorial search problem at scale. This is not a fine-tune-GPT-4 situation. The research frontier and the product are the same thing. - **Data: 8/10**: Verified math corpora are among the scarcest, most expensive-to-produce datasets in existence. Building that corpus is Cajal's primary moat and also their primary engineering challenge. You can't scrape your way to this. - **Backend: 7/10**: Deep Lean integration, proof search algorithms, sub-millisecond RL environment latency. Real systems engineering with very little room for slop. Correctness is non-negotiable when your product is correctness. - **Frontend: 2/10**: It's early-stage B2B selling to AI researchers and enterprise engineers. The UI is probably a dashboard, an API key, and a Slack channel. Nobody is buying Cajal for the UX. - **DevOps: 4/10**: Standard cloud infrastructure. Proof checking is computationally intensive but not architecturally exotic. Nothing here that a senior SRE hasn't seen before. ## The Moat: What's Real and What Isn't The cynical read on Cajal is that Lean is open source, AlphaProof proved the approach works, and a well-funded team could replicate the architecture. That's true. The architecture is not secret. But the moat isn't the architecture. It's three things that are genuinely hard to clone: **The data corpus.** Years of curated, formally verified mathematics across multiple proof assistants. Every theorem in that corpus was either written by hand by someone who knows what they're doing, or generated and verified by a system that has already been trained. You can't buy this data. You can't generate it without already having it. This is the most defensible asset Cajal has, and it compounds over time. **The expertise overlap.** You need people who are simultaneously strong in formal methods (dependent type theory, Lean's metaprogramming, tactic engines) and modern ML (multi-agent systems, RL from formal feedback, fine-tuning on proof corpora). That Venn diagram is tiny. Axiom raised $200M and is still hiring. DeepMind has an entire research team on this. The talent constraint is real. **First-mover in RL environments.** If Cajal gets their RL environment product embedded in a frontier lab's training infrastructure, that's a switching cost. Training pipelines don't get replaced casually. The team that solves the integration problem first has a durable advantage. What's *not* a moat: the model architecture, the general approach, the use of MCTS in proof search. These are known. Well-resourced competitors (big labs and funded startups alike) will get there. The honest summary is that Cajal's moat is time-based, not structural. They need to get far enough ahead that catching up becomes economically irrational. That's a race, and they're in it. ## Replicability Score: 72/100 This is a 72, not an 85, because the underlying approach is proven and the tools are open source. A strong ML team with formal methods expertise and serious funding could reproduce the core system. AlphaProof showed the world the recipe. Lean and Coq aren't proprietary. The architecture of Tau, multi-agent proof search with formal verification as the reward signal, is the kind of thing that gets written up in papers. It's not a 55, because the data moat is real and the expertise requirement is brutal. You're not building this with generalist engineers. The formal verification side alone requires people who have spent years thinking about type theory and proof assistants. The ML side requires people who understand why training on formal proofs is different from training on natural language. The combination is rare enough that it functions as a real barrier. The 28 points of difficulty that keep this from being fully replicable are concentrated in two places: the corpus of verified mathematics they're building, and the relationships they're establishing with frontier AI labs right now. Both of those compound. Both of those are hard to fast-follow. Cajal is playing in a space where the largest AI labs in the world are paying attention. That's both validation and threat. The question is whether a two-person team can move fast enough and build deep enough customer relationships to become infrastructure rather than competition. YC's Diana Hu backing them suggests at least one smart person thinks the answer is yes. Watch this one. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory.