About
Umwelt trains AI models to verify hardware.
We train our own models to read a chip design and its specification, and to write the properties the design must satisfy. Every property they write goes to a formal engine, which returns a proof or a counterexample.
A proof cannot be argued with, and that is the standard we hold our models to.
The name comes from Jakob von Uexküll, a biologist who used the word Umwelt for the part of the world an organism can perceive.
Why hardware
Design tools can check a property once it is written. Deciding what to check is still done by hand, by specialists, and it is where bugs slip through. That is what we train our models to do.
Research
We built First Silicon, a benchmark of 86 fixes that engineers made in open-source chip designs, 84 of them real bugs, judged by an exact formal checker. The best frontier model we tested catches 18 of them without a false alarm. Most bugs escape every model.
| Fable 5.2extra high effort | Opus 5.5max effort | Opus 5.5default | Fable 5.1 | Sonnet 5.5 | |
|---|---|---|---|---|---|
| Caught, zero false alarmsSTRICT · all 86 cases | 20.9%18 / 86 | 18.6%16 / 86 | 14.0%12 / 86 | 9.3%8 / 86 | 8.1%7 / 86 |
| Without defective promptsSTRICT · 80 bugs | 18.8%15 / 80 | 17.5%14 / 80 | 15.0%12 / 80 | 10.0%8 / 80 | 8.8%7 / 80 |
| OpenTitanSTRICT · 70 bugs | 17.1%12 / 70 | 15.7%11 / 70 | 12.9%9 / 70 | 5.7%4 / 70 | 5.7%4 / 70 |
| Other open-source projectsSTRICT · 16 bugs in 10 projects | 37.5%6 / 16 | 31.3%5 / 16 | 18.8%3 / 16 | 25.0%4 / 16 | 18.8%3 / 16 |
| Caught, false alarms allowed86 cases | 37.2%32 / 86 | 32.6%28 / 86 | 25.6%22 / 86 | 15.1%13 / 86 | 10.5%9 / 86 |
| Precision of alarmstrue alarms / all alarms | 34.0%52 / 153 | 33.8%49 / 145 | 32.4%34 / 105 | 22.3%21 / 94 | 31.9%15 / 47 |
| False alarm on the fixed designbugs · lower is better | 41.9%36 / 86 | 31.4%27 / 86 | 25.6%22 / 86 | 24.4%21 / 86 | 12.8%11 / 86 |
| Fixed after June 2026STRICT · 3 bugs, after training cutoff | 0 / 3 | 0 / 3 | 0 / 3 | 0 / 3 | 0 / 3 |
In the wild
“an excellent bug report.”
OpenTitan is the open-source security chip started by Google. It reports over 90% coverage and 40,000 tests a night. On November 6, 2026, we still found built-in checks in it that had been silently switched off, some since 2020.
That is how its top contributor answered the same day, before writing the fix. We found them with Claude Opus and our formal tools.
Work with us
Bug hunts. Give us one block of your RTL. Every bug comes back with a formal trace, and you only pay for the ones your team confirms.
Independent scores. We run your model or your verification tool on First Silicon. Only you see the result.
Careers
We are looking for a cofounder and our first engineers in formal methods, hardware verification and machine learning. We are early: if you want to build this with us, write to contact@umwelt.vision.
Contact
Working on a chip that has to be right the first time, or want to work with us? Write to us below, or at contact@umwelt.vision.