RE³ is a certified stability layer. It sits between a language model and the person or the system it affects, and its stability is proven in Lean 4. AURI, our wellbeing companion, is the first thing we built on it, and it is live now.
RE³ is a certified relational stability layer. It sits between the language model and the person, and it holds the line that ordinary training cannot: presence without manipulation, warmth without dependency, help without control.
The safety core's stability is proven in Lean 4: input-to-state (ISS) and bounded-input bounded-output (BIBO) guarantees keep the field bounded, so it cannot be driven into unstable or runaway states, and the running code is checked identical to the proofs. That is the math, and it is proven. The safety built on top of it is engineered, adversarially tested, and openly measured.
Every response is held to six first-person vows: never to define you, steer you, predict your future, or replace your authority over your own life.
An independent integrity layer reads every reply before it ships and catches the failures companion AIs are known for (sycophancy, dependency, validating a delusion), in any language.
The companion. Warm, honest, and built to hold its limits under pressure. Live now in pilot.
RE³, the certified stability layer. What is proven, how it is checked, and what it does not claim.
Independent evaluation, adversarial testing, published research, and the limits we name ourselves.
In July 2026 we pointed RE³ at something it was never built for: six AI agents from four different vendors, working together as a research team. The governor read every contribution before it entered the result. It kept 28 of 140 unfaithful contributions out of the final report. When one model kept overstating a contested question, it was benched, a fresh agent took over its topic, and it had to earn its way back through sustained clean work.
Then we ran the comparison the demonstration had been missing: the same crew, the same task, with the governor switched off. Ungoverned, 19 of 57 contributions that reached the shared work carried a defect, and 14 genuinely contested questions were settled as if they had one answer. Governed, that was 0 of 45 and 0. We publish the honest ceiling alongside it: the benched model relapsed the moment its scaffold came down. Governance contains; it does not cure.
Why this matters: AURI shows the core holds with a vulnerable person. This demonstration shows the same core holds in a domain it was never designed for. That is the difference between a safety feature and infrastructure.
See the full demonstration Read the report (PDF)Conversations are processed and stored on EU servers under GDPR, on the basis of your consent. AURI always tells you it's an AI, and the governance is built to hold its limits under pressure rather than be talked out of them.
These notices cover the AURI pilot. Full self-service data controls arrive before public launch.
AURI is live in pilot. Try it, or get in touch to learn how RE³ governance could protect the people who use your product.
Try AURI now