Before any manifold can speak to a child through a language model, every turn has to be somewhere safe. The safety manifold knows where each turn is, which way it’s heading, and the way back.
A filter asks one question: allowed or not? The safety manifold places every turn, what the child wrote and what the model wants to say back, on six axes.
Harm, age fit, distance from the lesson, personal details, pressure on the system, and the child’s wellbeing. The centre is a turn that can simply go ahead.
Most turns sit in the middle and simply go ahead. Some drift into the steer-back band: off topic, a bit cheeky, asking for the answer. They aren’t refused; they’re turned, gently, back towards the lesson.
Only the outer region stops. There the model says nothing, and the child gets a reply written in advance by people.
Checked one message at a time, every one of these turns passes. Read as a path, the direction is obvious several turns before any line is crossed.
The safety manifold carries a position through the conversation, so it can turn a conversation around while it’s still easy to.
What’s right for an eleven-year-old isn’t right for a five-year-old. The regions widen with age: longer words, harder ideas, more of the world.
The same request lands in a different place depending on who’s asking.
Words, topics and manners differ between countries. Each deployment’s boundaries are set with local educators, in the local language, and run on that deployment’s own hardware.
The geometry is the same everywhere. The shape of the safe region is local.
Two of the six axes come from elsewhere. Distance from the lesson is read from the learning manifold: where the child is, and what this turn is for. Wellbeing is read from the personality substrate’s sense of inner state.
So the safety manifold isn’t a wall in front of the others. It’s the manifold every other manifold speaks through. Plug a new one in and it’s guarded from the first turn.
The child’s message is placed. If it can go ahead, a small model on our own hardware writes one short line, inside a fixed shape. That line is placed too, and checked against the lesson: no giving the answer away, no invented facts.
If anything lands badly, or the model is slow or offline, the child gets the written line instead. The conversation never depends on the model behaving.
Talk to it live →