They check the model. We govern the change.

We do not test robot models. Bodies such as the Physical AI Safety Institute ask whether a robot foundation model is safe: they probe it, benchmark it and find the ways it can be broken. Our question starts where theirs ends. A robot is approved once; then its model changes. We work on what happens after that: how conformity assessment survives a model update, how one model running on many bodies fails at fleet scale, and how fast machines are coming to build machines.

We are the customers and translators of technical safety work, turning findings into text that standards committees and regulators can adopt.

Robots are now certified once and then updated continuously; soon they will be built by other robots. We keep the public record of physical-AI incidents with a robot category and a model-version field together, build the indicators and thresholds that let regulators and standards bodies see the physical world the way they already see compute, and write the clauses they can adopt.