&Truth launches to automate mathematical checks of software with AI

&Truth publicly launched on October 29, 2026, aiming to use AI to automate formal verification: mathematical checks that software behaves as intended. The startup is targeting a process whose cost and complexity have limited adoption, with research from a scientist who joined the company pointing to sharply faster work.

A research result behind the launch

Stanford researcher Chuyue Sun joined &Truth in August. In research described by WIRED, Sun and fellow computer scientists developed an AI tool that completed a verification engineering project in less than half a day for under $500. Three highly skilled experts had taken five months to do the same work manually.

The work used AI to verify part of the code that encrypts conversations in Signal. That makes it a concrete example involving security-sensitive software, rather than a claim that AI has proved an entire messaging app free of bugs. The result came from Sun’s research, not a disclosed test of a finished &Truth product.

&Truth’s ambition reaches further: it aims eventually to verify software in any programming language. That is a goal, not a capability established by the launch. AI-assisted verification remains at the prototype stage, and customers’ willingness to pay and preferred ways of adopting it are still uncertain.

A proof starts with a promise

Formal verification requires engineers to state precisely what software should do, then convert code into mathematical formulas that can be checked. Conventional testing looks for failures by exercising software. Verification can catch behavior those tests miss, but producing the statements and proofs has traditionally demanded substantial specialist work.

The automation effort targets that preparation. Startups working in the field generally plan to use AI to help define a program’s required behavior. For a sorting operation, for example, the rule might require a table to contain the same items before and after sorting, with only their order changed.

Further automation would translate code into a form that mathematical tools such as Lean, Rocq and Isabelle can check. That translation need not itself use AI. The proposed workflow combines help writing requirements with machinery for checking whether code satisfies them; it is not simply another model reviewing code and declaring it correct.

The technology to mathematically reason was all there, just difficult to scale.

Ilya Sergey, &Truth chief science officer, speaking to WIRED

Stronger guarantees, bounded by assumptions

The strongest competing interpretation concerns the assurance a proof provides. Formal verification depends on assumptions about how computers and people interact with software. Some experts question its ability to catch failures caused by user error or hardware breakdowns. A mathematical result therefore does not settle every question about how a system will behave.

Will Wilson, who runs bug-hunting software startup Antithesis, offers a practical objection. His company spent millions formally verifying some cybersecurity defenses, yet later found additional errors with its own testing tool. He told WIRED that Antithesis has also found bugs in software its customers had formally verified.

There is also a risk at the requirements stage. Proofcraft CEO June Andronick and chief scientist Gerwin Klein worry that AI-generated specifications could create unjustified trust in verification results. Their concern sits directly beside the automation promise: making requirements easier to produce does not, by itself, establish that those requirements deserve confidence.

The commercial hurdle is equally concrete. Midsized technology companies told WIRED that scarce verification engineers, costs and weeks-long reviews have restricted their use of the method. &Truth is trying to automate work that companies have often judged too expensive or slow. Sun’s result addresses that burden; broader adoption remains unproven.