How Atlas decides what your child does next.
Every learning app has content. The hard part is choosing. This is the part of Atlas that chooses — it runs on the phone, it uses no AI model at runtime, and it can explain every decision it makes.
Atlas decides what a child does next with a deterministic engine running on the phone: it holds a probability that the child knows each skill, walks a map of which skills depend on which, schedules review before the memory would fail, and uses no AI model at runtime.
A probability, not a score.
For every skill, Atlas holds a probability that your child knows it — updated on every single answer.
It is not a running percentage of right answers. It explicitly models guessing (a child can get a three-choice question right by luck) and slipping (a child who knows it can still tap the wrong thing). So a lucky streak does not read as knowledge, and one careless tap does not erase it.
The technique is Bayesian Knowledge Tracing, and it is about thirty years old. We did not invent it. We made it affordable.
A map of what depends on what.
Skills are arranged by what they require. A child who cannot yet partition ten is not shown two-digit addition — not because of their grade, but because the prerequisite is not there.
This is why Atlas can start a nine-year-old at the actual gap rather than at the top of Year 4. The map, not the calendar, decides.
It brings things back before they are forgotten.
Learning something once is not learning it. Atlas models the forgetting curve per skill, per child, and surfaces a review just before the memory would fail — not on a fixed weekly schedule, and not when your child happens to feel like it.
The scheduler is FSRS. Also not our invention. Also boring and proven, which is the point. Nobody in the house has to remember the review, and in many houses nobody could.
Four dimensions, never one number.
| The question it answers | |
|---|---|
| Independence | Did they do it without help, or with hints? |
| Retention | Do they still know it after a delay? |
| Transfer | Can they do it on a question they have never seen? |
| Evidence quality | How much do we trust this particular observation? |
Two things that forces us to do.
Atlas never reports a skill as mastered on same-session evidence. Answering right twice in five minutes is not knowing.
And a fully-revealed answer updates the model as though the child had got it wrong, because being shown is not evidence of knowing.
Why does Atlas aim for your child to be right only 85% of the time?
Because that is where learning is fastest. Formal models of learning put the most efficient training point at roughly 85% success — hard enough to be worth doing, easy enough to keep going. Not 100%, and not 50%.
This is why Atlas bans streaks, hearts and lives. At the efficient edge, a child is wrong about one time in seven and that is the tutor working correctly. Any reward that resets on a wrong answer would be draining exactly when the engine is doing its job.
Does an AI model run on my child's phone?
No. Generation happens at build time, on real machines, where a human reads the output before a child sees it. What ships to the phone is a deterministic engine: same input, same output, always.
It cannot hallucinate. It cannot drift. It can be audited. And it will re-teach the same skill the two-hundredth time exactly as patiently as the first.
Language models build almost everything Atlas ships. They just do not get to improvise at a child.
And this is rarer than it should be.
Kitkit School won the Global Learning XPRIZE and was validated in a randomised trial. We read its source. It has no adaptive engine at all — a fixed 529-day ladder, the same order for every child, no mastery model and no prerequisite map.
That is not a criticism. They authored the sequence because there was nothing that could compute it. The best-evidenced products in this category are hand-sequenced, one order for everyone.
That is the thing Atlas has.
An engine that can explain every decision it makes.
One email when Atlas opens. Nothing else.