What you get when your model writes LMAT.
Six things that hold for every program, whoever or whatever wrote it.
In this section
- Checked before it runs: 202 kinds of mistake, each named with the line and the fix.
- It tests itself: Every button pressed and every form filled, as every user.
- The same answer everywhere: It runs the same on every machine.
- It can’t reach what it shouldn’t: No other customer’s data, no web addresses, nothing that runs forever.
- Private by design: Each person sees only what they are allowed to.
- No escape hatch: There is no way around the checks.
Part of The first language designed for AI to write..
The whole tree
- The first language designed for AI to write.: Your model writes LMAT. A compiler checks every line before it runs.
- AI writes code faster than anyone can review it.: Review is the new bottleneck.
- Twice the code, twice the review: Measured across more than 10,000 developers by Faros AI.
- Checked before it runs.: The model writes, the compiler checks, it runs.
- One request, start to finish: A request, a refusal in the checker’s own words, a repair.
- Real software, built on LMAT.: LMAT makes them work. The design is free.
- Software: Accounting, an issue tracker and a design tool.
- Games: Rocket-car soccer and a block world.
- What you get when your model writes LMAT.: Six things that hold for every program.
- Checked before it runs: 202 kinds of mistake, each named with the line and the fix.
- It tests itself: Every button pressed and every form filled, as every user.
- The same answer everywhere: It runs the same on every machine.
- It can’t reach what it shouldn’t: No other customer’s data, no web addresses, nothing that runs forever.
- Private by design: Each person sees only what they are allowed to.
- No escape hatch: There is no way around the checks.
- Built for agents too.: lmatfs, the storage layer for AI agents.
- lmatfs: Rust, with Python bindings.
- Measured, not claimed.: Counted from the LMAT repository on 5 October 2026.
- 202 kinds of mistake caught: Each named with the line, the field and the fix.
- 5,635 tests passing: Across the engine, the sandbox and the renderer.
- 36 attack programs run before every merge: Programs that try to break out of the sandbox or steal from it.
- Same requests, before and after LMAT tested each app: 23 requests, September 2026: three broke, then none.
- Who is building it: Two founders out of Carnegie Mellon.
- The team: Uzi Isman and Jason Shen.
- Talk to us.: We’re raising, and we work with AI companies.