Bhavesh Shaha

Our autonomous prover Dirac proved all 6 IMO 2026 problems in 7h 18m, with machine-checked proofs. Hi everyone, I wanted to share something @13tcluelessbird and I have been working on for some time: Boundless Intuition @bi_labs. It started with an observation during our work at CERN: in scientific and mission-critical settings, an answer that sounds right isn’t enough. We need a way to verify that it’s actually correct. That led us deep into formal verification and a bigger question: What if we could make AI’s most consequential decisions provably correct, rather than simply trust its confidence? Dirac is one step toward that vision. We ran it against the publicly released Lean formalizations of all six IMO 2026 problems. Among the publicly reported systems we’ve compared against on the same formalizations, Dirac was the fastest. And mathematics is just one aspect. We’ve also published benchmarks exploring formal verification across aviation rules, clinical calculations, and tax legislation, including substantial improvements in accuracy and inference efficiency. The broader goal of study is to close the loop between AI-driven research and formal verification, enabling autonomous systems to conduct research in complex fields like particle physics, biology, manufacturing, and quantum science while verifying their results and actions. This could accelerate scientific discovery by allowing AI to explore, experiment, and improve autonomously, without sacrificing trust in the process. We’re sharing the work, benchmarks, and ideas publicly. → Overview & benchmarks: boundlessintuition.com/overv… → Dirac’s IMO 2026 results: boundlessintuition.com/blog/… → Towards Verified Superintelligence: boundlessintuition.com/blog/…