Why we think this teaches.
The evidence we built on, what it does and does not say, and the decisions it forced.
Atlas is built on published education research — mastery before pace, spaced retrieval, and a difficulty that targets roughly 85% success — and those studies belong to the people who ran them: they inform how Atlas was built, they are not studies of Atlas.
The studies below inform how Atlas was built. They are not studies of Atlas, and they do not guarantee any academic result for any child. Each one belongs to its authors, and we name them.
Anyone who cites learning science without that sentence is borrowing credibility they have not earned.
What does "mastery before pace" mean?
It means Atlas will not advance past a skill on a shaky estimate — whatever grade the child is in, and whatever week it is.
A child who has not understood something gains nothing from moving on: what comes next rests on exactly the piece they are missing, and every later step gets quietly harder for a reason nobody can see any more. This is the oldest idea in the field and the hardest one to sell to a calendar, because a calendar advances whether the child did or not.
Why does Atlas make children recall instead of review?
Because retrieving something strengthens it far more than re-reading it, and spacing those retrievals beats massing them together. It is the most robust finding in the study of memory.
So retrieval is the default activity in Atlas rather than an occasional quiz, and the spacing is computed per child and per skill rather than by the week. Recall is effortful, and that is why it works.
Why does Atlas let my child get things wrong?
Because the most efficient point to train is not getting everything right — it is roughly 85% success. Wilson et al. (2019) derive that band from formal models of learning, and Atlas's next-item policy targets it deliberately.
Hard enough to be worth doing, easy enough to keep going. It also explains a design decision that surprises people: at that rate a child is wrong about one time in seven, on purpose. Any mechanic that punishes a wrong answer is therefore punishing the tutor for working correctly.
Motivation: what is banned, and the rule underneath it.
No streaks. No hearts. No lives. No leaderboards. No timers. No confetti. This is enforced in our test suite, not merely intended.
The rule is not "no game feel." Children love stakes, collections and unlocks, and a product that strips them out is not virtuous, it is dull. The rule is about who controls the input, and a mechanic has to pass two tests.
1 · Is its input whether the child showed up? A day-streak's input is whether a parent handed over the phone on Tuesday. That is a fact about a family's week, not about a child's effort — and it punishes them hardest on exactly the days they most need a reason to come back.
2 · Does it reset on a wrong answer? See above: at the efficient edge, being wrong is correct operation.
| The mechanic | What decides the outcome | Verdict |
|---|---|---|
| Mario's lives | Your own jump and your own timing — the player controls it. | Passes both tests. |
| A day-streak | Whether the family phone was free on Tuesday. | Fails the first. |
| Hearts in a tutor | Whether the child already knew the answer walking in. | Fails the second. |
Stakes are fine. Loss is not. A child may chase something; it may not be taken away for a reason that was never theirs.
And we should say that this is a real disagreement and not settled science. Bloomy's own family page advertises "Bloomy Bucks" and a streak counter, and their founder defends it — "some amount of extrinsic motivation can cultivate intrinsic motivation." That is an honest position. We went the other way deliberately.
Can AI tutoring make learning worse?
Yes, and that is the finding that shaped this product most. In a study of about a thousand students (Bastani et al., PNAS 2025), students given unguarded access to a general AI assistant performed 48% better while they had it — and 17% worse than the control group on the exam once it was taken away. The version with guardrails did not harm learning.
The guardrails are the product. A tutor that makes a child look capable while it is present and less capable once it is gone has taught dependence, not mathematics. It is why Atlas measures unaided performance separately from everything else, and why it claims nothing on the strength of a session where it did the work.
The measurement is the product.
Most education software reports usage: minutes, lessons opened, days in a row. Atlas reports what a child can do without help, traced back to the evidence for it.
And some weeks it will say that not much moved. A report that can only ever show progress is not measuring anything.
Write to us if you want to argue with any of these decisions.
One email when Atlas opens. Nothing else.