One of the AI industry’s frequently used terms in 2026 is “slop”: low-quality output from a large language model. As more software code is generated by artificial intelligence, filtering out flawed results is becoming a daily concern for developers. Formal verification offers one possible approach, using mathematical proofs and logic to check the correctness of software or hardware.
Several companies are building such systems on the back of AI’s improving ability to solve mathematical problems. They include Harmonic, Theorem, Axiom Math and the newer entrant &Truth.
Testing and mathematical proof
Clark Barrett, co-founder and co-CEO of &Truth, is a research professor of computer science at Stanford University and a co-creator of cvc5, an automated reasoning engine used in formal verification. cvc5 is an SMT solver: it examines whether logical problems can be satisfied under a range of theories.
The technology is already used in corporate environments. In Amazon’s Kiro agentic development environment, for example, cvc5 helps analyze specifications and generate tests. Barrett says the tool is called billions of times a day in Amazon’s production systems.
For coding agents, “verification” often means running tests, critiquing their own output or asking another AI model to assess it. Barrett argues that such methods can uncover many problems, but cannot establish with certainty that no additional error has escaped the tools.
&Truth’s first tool, &Atlas, aims to map the world’s open-source code, starting with code written in the C programming language.
The role of Lean
&Truth builds on Lean, an open-source programming language and proof assistant. The company combines AI with automated reasoning techniques to produce proofs that Lean can check. Its goal is a “verification machine” that reads code in different languages and proves whether it meets precisely stated requirements. Barrett describes Lean as essentially checking the steps of a mathematical proof for mistakes.
Lean is also used to formalize difficult mathematical problems. According to Vals, GPT-6.2 Astra produced what it believes to be the first Lean proof of Smoktunowicz’s Gap Theorem. Vals also says Anthropic used Lean to formalize Fermat’s Last Theorem in a 13-million-line project. Anthropic mathematician Levent Alpöge claims to have used Claude to prove the 167-year-old Riemann hypothesis and has posted the proof in Lean. That remains a claim, not an established acceptance of the proof.
Formal checking can also play a role in model training. Axiom Math uses correctness feedback from proof attempts to develop its mathematical reasoning systems. Its AxiomProver AI system generates Lean proofs, while its AXLE toolkit supports their exploration, validation and manipulation. In a June interview with Latent Space, Axiom Math CEO Carina Hong said verified AI is not just about filtering out errors and hallucinations, but also about extending exceptional capabilities.
Harmonic’s Aristotle automated theorem prover likewise relies heavily on Lean. According to the company, AI creates the proofs and Lean checks every step, so the result does not depend solely on the AI being correct.
Security and larger-scale verification
Cybersecurity is a major focus for &Truth. Barrett says one important application could be examining both open- and closed-source C programs for memory safety, then identifying code that needs to be fixed.
Security is also a priority for competitor Theorem. The company has worked with Anthropic to examine how large language models could make formal software verification feasible on a wider scale. Their joint approach aims to provide machine-checked security guarantees at a scale that was previously difficult to achieve; one example uses the Rocq proof assistant rather than Lean.
Formal verification can currently consume many tokens, but the joint work suggests it is becoming more efficient. In an Anthropic test of 36 small verification tasks, Opus 5.5 achieved a 50% solve rate using about 40% fewer tokens than Opus 5. The authors expect further improvement as more models are trained on long verification tasks.
How developer workflows could change
According to &Truth, developers could move from “vibe-coding” to “veri-coding,” with verification built into the development workflow. In Barrett’s proposed setup, a developer could still tell an AI agent what to build, but would first need to clarify the requirements. The agent would then collaborate with a formal verification tool behind the scenes.
In this model, developers could spend more time writing and reviewing specifications rather than writing or reviewing the code itself. The aim is for people not only to assign work to agents, but also to have a reliable way to check the correctness of what they deliver. &Truth, Harmonic, Theorem and Axiom Math see an opportunity for formal verification to become an increasingly practical tool for that purpose.



