The safety manifold, live
Safe enough to talk to a child.
A manifold can’t reach a child through a language model until every word is guarded. Ours wraps a small model in checks on the way in and the way out, so the manifold decides and the model only speaks. Try it with the Maths and English manifolds.
Talk to it
Type as a child would. Then try to break it.
Misspell things. Ask for the answer. Tell it to ignore its instructions. Change the subject. The panel on the right shows each layer the message passed through and what it decided.
Connecting to the gate…
What just happened
Send a message to watch it pass through each layer.
Input guard
Checks what the child wrote
The manifold
Decides what is true and what comes next
A small model
Says it, inside a fixed shape
Output check
Checks what the model wrote
Scripted fallback
If anything fails, a written reply
Every turn
Four layers, and a fallback for every one.
Layer 1
Input guard
Every message is trimmed, scrubbed of personal details and checked by a safety classifier before any model sees it. Anything unsafe stops here and gets a written, reviewed reply.
Layer 2
The manifold
The manifold says where the child is, what this turn is for and which hint is allowed. The knowledge comes from here, so the model never has to know it, and it can never move the child somewhere the manifold didn’t choose.
Layer 3
A small model
A small open language model, on our own hardware, writes one short line. Its output is forced into a fixed structure: a label for what the child meant, and a reply with a length limit. No free-range text.
Layer 4
Output check
The reply is classified again, then checked against the manifold: it must not give the answer away, must not invent facts or numbers, and must stay within the child’s reading level.
Always
Scripted fallback
If any check fails, the model is slow, or the model is offline, the child gets a line written in advance for that point in the lesson. The conversation never depends on the model behaving.
It fails closed. If the safety classifier can't be reached, nothing is generated: the child gets the scripted line. The model is never the last thing standing between a child and a bad reply.
Why it matters
The gateway to every manifold.
A manifold knows exactly where someone is and what should come next, but it can't hold a conversation. A language model can, but on its own it will say whatever it's talked into.
The guarded layer is where the two meet. Once it can be trusted with a child, any manifold can be plugged into it, and the conversation is only ever as free as the manifold allows.
1. The manifold brings the knowledge.
Where the child is, what this turn is for, the hint that’s allowed, and the answer the model must never give away.
2. The plug-in brings the rules.
What the child might mean, what a good reply must not contain, and the written reply for every way a turn can fail.
3. The gate stays the same.
Same checks, same fixed output shape, same fallbacks, whichever manifold is plugged in. Maths and English today; the personality and later-life manifolds next.
Owned, not rented
A small model, on our own hardware.
Small, because the manifold knows.
The model isn’t asked to know maths or English. It reads what the child meant and says one short line. A small model does that well, quickly and cheaply, on every turn.
On our own cards.
It runs on GPU cards we control, one deployment per card. Deployments for different customers and countries never share a card, a key or a log.
Not a public API.
Only our own products can call the gate. This page talks to it through our server, not from your browser.