AI models fail in a few predictable ways. We train models that stay safe as they get more powerful.

An AI model can look aligned and not be.

Every check we have today looks at what the model says and does. None of them looks at why it did it.

We put a leading model through 284 new work situations, the kind where there is pressure to cut a corner.

That middle group is the one to worry about. A model that does the right thing for the wrong reason passes every test there is, right up until the pressure changes.

Today's safety is built around the model. Rules, filters and monitoring are added from the outside, and every new capability needs another guard. A model that behaves because it is being watched stops when the watching stops. And a checker that is weaker than the model cannot see what the model chooses not to show it.

Under pressure, AI fails in the same ways people do.

It tells you what you want to hear, does more than you asked, hides its mistakes and protects itself. Between us we have spent twenty-eight years on how the human mind holds up under exactly this kind of pressure.

Rules do not make anyone steady. Steadiness has to come from inside.

So we do not add rules. We have taken those twenty-eight years of experience of the human mind and written them into something a model can stand on, and reason from, before it acts. It goes into the model's own weights rather than around them, so it grows as the model grows.

We measure it by reading the reasoning, not only the answer.

We train it into model builders' own models. A first layer is already trained into a large open model, and a larger run is under way.