Theorem raises $6M to automate mathematical checks of AI-written software

A startup emerging from YC has closed a $6 million seed round to commercialize an automated formal-verification pipeline aimed at AI-generated code. Backed by a mix of institutional and angel investors, the company layers machine-learned models on top of mathematical proof techniques to accelerate tasks that once required elite specialists. The core technical claim is compressing verification workflows that historically took years into processes that run in weeks or days by automating proof construction and management. In internal demonstrations the platform translated more than a thousand formal problems between proof languages and produced automated equivalence checks that would otherwise consume multiple person-years of labor. A customer case showed the system can reimplement a complex specification into a production parser and raise throughput from low-megabit to gigabit rates while providing a formal harness that asserts behavioral equivalence. Architecturally, the product splits verification obligations into smaller units so agents can execute in parallel or in sequence, allowing it to handle interdependent code that defeats context-limited assistants. The team emphasizes defects found by formal proofs that slip past conventional testing, arguing that test suites alone are an incomplete assurance model as AI accelerates code churn. Founders combine deep programming-language and verified-systems experience with machine-learning engineering to train models that generate and check proofs at scale. The company plans to add staff, scale compute for model training, and push into robotics, renewables, crypto, and drug-research toolchains. If the automation reliably reduces human effort and integrates with engineering workflows, it could change how enterprises and regulators assess the safety of AI-produced software. If adoption stalls or the models make brittle assumptions, the benefits will remain confined to specialized, safety-critical projects.