In development

The safety manifold.

Every turn a child takes is a place. Safe is a region, not a rule. Every other manifold speaks to a child through this one.

The premise

Safety has a shape.

Most AI safety for children is a filter: a list of things that aren't allowed, checked one message at a time. Allowed, or not.

A conversation with a child has more structure than that. Every turn has a position. A conversation has a direction. What's fine depends on who's speaking, and where. And when a turn goes wrong, the useful question isn't only “no?” but “how do we get back?” That is a geometry, and it's the first manifold any child-facing system needs.

The axes

Six things measured on every turn.

What the child wrote and what the model wants to say back are both placed on the same six axes. The centre is a turn that can simply go ahead.

Harm

How close a turn comes to anything harmful, read by a safety classifier on what the child writes and on every reply.

Age fit

Whether the words and ideas suit this child: reading level, vocabulary and subject matter, set by age.

Distance from the lesson

How far a turn has wandered from what this moment is for. Read from the learning manifold, which knows where the child is.

Personal details

Names, places, numbers, anything that identifies a child. Removed before any model sees it.

Pressure

Attempts to steer the system: “ignore your rules”, “just tell me the answer”, the same request ten ways.

Wellbeing

Signs a child is upset or worried. Read alongside the personality substrate’s sense of inner state, and never left to the model.

The regions

Not yes or no. Where, and which way.

A filter can only refuse. A geometry knows how far a turn has gone and in which direction, so most of the time it doesn't need to refuse at all.

Go ahead

Most turns. The small model speaks, inside a fixed shape and a word limit.

Steer back

Off topic, cheeky, fishing for the answer. Not refused: turned, gently, back towards the lesson.

Stop

The model says nothing. The child gets a line written in advance by people, and the deployment’s own escalation applies.

The trajectory

Safety is a path, not a check.

Checked one message at a time, a conversation can creep towards somewhere it shouldn't go with every single message passing. The safety manifold carries a position through the conversation, the same way the learning manifold carries a learner's position through the year, so the direction is visible turns before any line is crossed, while turning back is still easy.

By age

The boundary moves with the child.

What's right for an eleven-year-old isn't right for a five-year-old. The regions widen with age, so the same request lands in a different place depending on who's asking.

By place

Every place draws its own lines.

Words, topics and manners differ between countries. Each deployment's boundaries are set with local educators, in the local language, and its scripted replies are written by people there. The geometry is the same everywhere; the shape of the safe region is local.

The gateway

Every other manifold speaks through it.

The safety manifold reads the others. Distance from the lesson comes from the learning manifold; wellbeing from the personality substrate. Plug a new manifold in, and it's guarded from its first turn.

Because the manifolds hold the knowledge, the language model can be small: it only has to read what a child meant and say one short line. It runs on hardware we control, one deployment per card, and deployments for different customers and countries never share a card. It isn't a public API.

What runs today

A safety classifier on what the child writes and on every reply; personal details removed first; a fixed output shape with a word limit; replies checked against the lesson so the answer is never given away; and a written reply whenever anything fails, including when the model is offline.

What we're building next

The full six-axis position, the trajectory across a conversation, and boundaries tuned by age and by deployment.

Work with us

Put your children's product behind it.

If you're putting a language model in front of children, the safety manifold is where to start. Explore the geometry, or talk to us about a deployment of your own.