The Seven Safety Layers, Explained for Parents
The guard at the gate, the form teacher who reads the question first, the teacher, a check before anything is said aloud, and a note home. Our safety pipeline, told in the language of a school day.
Walking Your Child to the Gate
When you hand your child to a school each morning, you are not trusting one person. You are trusting a system: the guard at the gate, the form teacher, the subject teacher, the discipline master, the note that comes home. No single adult carries the whole weight, so no single adult's bad day becomes your child's bad day.
We built OpenKids the same way. Every message your child sends passes through layered, independent checks — before the AI, inside the AI, and after the AI — and each layer assumes the others might miss something. Here is the walk through our school day, layer by layer, in plain language.
The Guard at the Gate: Crisis Interception
Before any message reaches any AI, it is screened for signals of real distress — self-harm, abuse, danger. If such a signal is found, the conversation does not proceed to a generative model at all.
What your child sees instead is a fixed, human-written response: warm, calm, and pointing to real help. Not generated, not improvised — the same input always produces the same safe words, because a crisis is the last place a machine should be creative. At the same moment, an email goes to you. Not in the weekly report. Immediately.
The Form Teacher Reads the Question First: Input Checks
Next, the message meets two readers before any coach does. First, a set of local rules — fast, free, deterministic — that stop the obviously out-of-bounds on the spot. Second, a separate AI whose only job is moderation reads the message semantically: not keyword matching, but meaning. If your child attached a photo, the photo is screened right alongside the text.
One detail we are particularly stubborn about: if that semantic reviewer ever fails — an outage, an error — we block the message rather than wave it through. Engineers call this failing closed. A jammed school gate should stay shut, not swing open.
The Teacher: Models Routed by Age
Only now does the question reach a coach — and which underlying model answers is not one-size-fits-all. Routing takes the child's age and the subject into account: a nine-year-old's question is handled differently from a seventeen-year-old's, in model choice, in tone, in how much is said at once.
And the teacher never walks in unprepared. Every coach carries its own persona, syllabus alignment and safety notes, plus the age-band instructions that tell it how to speak to this child. The freewheeling improvisation of a general chatbot is exactly what this layer is built to remove.
Before Speaking: The Reply Is Checked Twice
The coach's own words are not exempt. Before a reply reaches the screen, it goes through the same two-stage review the input did: local rules first, then the semantic moderator reads what the AI is about to say.
This is why our coaches do not stream words onto the screen the instant they are generated. The full reply is buffered, reviewed, and only then shown. It costs a beat of waiting, and we pay it gladly — slow is annoying, unsafe is unacceptable. Anything that fails review is replaced before your child ever sees it.
The Note Home — And Why None of This Is a Slogan
Every safety event, from a mild flag to a hard block, is logged with its risk level and shown in your parent dashboard, alongside the weekly learning report. A crisis event does not wait for any report: it emails you the moment it happens. Even our own staff reviewing a child's messages leaves a trace — each such view is written to an audit log.
Here is the part we most want parents to hear: everything above is not a marketing diagram. It is code, and it ships with automated tests — crisis inputs that must trigger the fixed response, blocked content that must stay blocked, fail-closed behaviour that must actually close. The suite runs on every release. If a safety test fails, that release does not go out. That is the whole policy, and it has no exceptions.