How it works, and why
For each method we wrote what it does, why children need it, how strong the evidence behind it is, and what we deliberately do not do.
A note on scale
Today one founding family uses it every day. We make no claims yet about learning results. What we can say is that each method was chosen from published research, that we know how strong that research is, and that we built the system to measure itself as it grows.
How to read the labels
- Strong
- Several controlled studies or meta-analyses point the same way.
- Moderate
- Good studies exist, but in other ages, settings or subjects than ours, or with mixed results.
- Limited
- Small, early or indirect evidence. We use it as a direction, not a promise.
- Not established
- No evidence we could find. Where we still do it, we say it is our design choice.
What we publish and what we do not
We publish the idea, the reason, the evidence and the limits of every method. We do not publish the exact numbers we use, the instructions we give the AI, how it is assembled, or the code. The numbers are starting points we keep tuning, and the instructions and code are the work itself.
Practice that adjusts itself
Status: Live in the prototype. A new part of the model is running in a side-by-side trial and does not decide anything yet.
What it is
For each skill we estimate how well the children know it, and for each kind of question how hard it is. From the two we predict the chance of a right answer, and after each answer both estimates move a little. It is an idea from chess, adapted for learning and used by large practice systems for children.
Why
Children do not have a single "level". They may be fluent at multiplication and new to division. This lets us pick questions they mostly get right, learn which questions are truly hard, and check whether our predictions hold.
What the research says
- Evidence: strong As an engineering method: such models predict well and work at scale (Pelánek 2016; Brinkhuis et al. 2018).
- Evidence: moderate For practice with a high success rate helping learning and motivation (field studies, not all at our ages; Brinkhuis et al. 2018; Papoušek & Pelánek 2015).
What we deliberately do not do
We do not aim for "half right, half wrong", which is the target for tests, not practice. No ticking clock, and a slow right answer still counts as right. The rating is not shown to children and never compares them.
Shared control over difficulty
Status: Built, not yet released
What it is
By default questions are "just right". At any moment the children can choose "a bit easier" or "a bit harder". The words describe the questions, never the children.
Why
Children are more motivated when they have a real say. But a choice that costs something is not a real choice, so it ships only together with fair rewards (see "Fair motivation").
What the research says
- Evidence: moderate Choice raises intrinsic motivation, effort and performance, and in a meta-analysis the effect was stronger for children (Patall et al. 2008; Ryan & Deci 2000).
- Evidence: not established That children left alone mostly pick easy work. That is our design assumption, and we will measure it.
What we deliberately do not do
No labels on the children, no reward or penalty for any setting, and the system never secretly overrides their choice: it may offer, never force.
Practice signals, without reading emotions
Status: Live in the prototype. Basic support is live; further parts are built or planned.
What it is
From answers alone we notice a few patterns: answering faster than anyone could read the question, getting stuck, questions that are too easy, or skipping. Each pattern leads to one gentle move: a hint, a different explanation, a step easier or harder, or an offer of a break they can decline.
Why
Children get stuck, guess and drift off. A good tutor notices and changes approach without making the child feel judged.
What the research says
- Evidence: moderate Rapid guessing is the best-validated sign of "not really trying" (Wise & Kong 2005), and a loop of failures signals that a different explanation is needed (Zhang et al. 2019). Most of this research is from tests, not practice.
- Evidence: not established For detecting emotions from on-screen behaviour: such detectors agree only weakly with human observers (Baker et al. 2012), so we do not use them.
What we deliberately do not do
No camera, no mood detection from the voice, no use of screen pressure or typing rhythm. The system never tells children what they feel. Breaks are offered, not forced, and saying no costs nothing.
Spaced review and two-step mastery
Status: Built, not yet released
What it is
Things to remember, like English words and maths facts, come back just before they would be forgotten. The guide offers a few as an optional "flashback" that can be skipped. A skill counts as "mastered" only after one more success on a different day.
Why
Children forget, and reviewing at the right moment is one of the most reliable ways to make learning last.
What the research says
- Evidence: strong For spacing in general (Cepeda et al. 2006; Dunlosky et al. 2013).
- Evidence: moderate For our age group: the effect is real but may fade sooner (Vlach & Sandhofer 2012; Barzagar Nazari & Ebersbach 2019; Lindsey et al. 2014).
What we deliberately do not do
No overdue counters, no red badges and no "your words are fading". A review not done simply waits. A missed later check never takes away a badge or a level.
A quiet corner and feelings tools
Status: Live in the prototype. The quiet corner is live; the feelings toolbox is approved for the founding family and not yet released.
What it is
At the top of the lighthouse island is a quiet corner with short tools: breathing with the lighthouse beam, shake and stretch, counting waves, squeeze and release, sea sounds and an evening story. It opens at any time, and one button takes them back where they were.
Why
Children need simple ways to settle their body and name a feeling, and small frustrations happen in a learning app.
What the research says
- Evidence: limited A meta-analysis of mindfulness for children found a small effect, and a large school trial found no benefit (Dunning et al. 2019; Kuyken et al. 2022). School social-emotional programmes helped (Durlak et al. 2011), but for a home app the effect is not established.
What we deliberately do not do
No therapy claims and no promise that a tool "will calm you down". Nothing measures calm: no heart rate, no breathing through the microphone, no face. Parents see which tools were used, never the feelings.
Fair motivation
Status: Built, not yet released. "Days together" instead of streaks, and coins without penalties.
What it is
Coins and small rewards reward trying and progress, never punish mistakes. Hints are free. A harder choice does not earn less. Guessing does not pay. "Days together" only go up, instead of a streak that resets.
Why
Rewards can quietly change why children do something. A streak that breaks turns a missed day into a loss, which is exactly the kind of pressure a children's product should not use.
What the research says
- Evidence: moderate Expected, tangible rewards can lower intrinsic interest while informational feedback raises it; the topic is disputed (Deci, Koestner & Ryan 1999; Cameron & Pierce 1994).
- Evidence: limited On streaks for children, with a clear regulatory direction (Silverman & Barasch 2023; ICO standard 5).
What we deliberately do not do
No coin loss, no paid hints, no surprise boxes, no leaderboards between siblings, and no badge or level is ever taken away.
A guide that never chases
Status: Live in the prototype. AI disclosure, break suggestions, the script for worrying moments, and an editable memory book.
What it is
Each child chooses a guide, names it and dresses it up. It speaks Hebrew, teaches English, remembers only what the children chose to share, and celebrates the work. It is written so it cannot pressure: it never says it missed them, never claims to be the only friend, points them toward people, and never makes first contact.
Why
Familiar characters can help children learn, and the same bond can be used to keep them playing. We want the first and never the second.
What the research says
- Evidence: limited For the benefit at our ages (Gola et al. 2013). The risk is well documented: characters pressuring children to keep playing were found in apps for young children (Radesky et al. 2022).
What we deliberately do not do
No guilt, no exclusivity, no notifications from the guide, no decay while away, and no praise of the child as a person, only of the work.
Learning by teaching the guide
Status: Built, not yet released
What it is
After a lesson they can choose to explain one idea to the guide, by voice or by ordering picture cards. The guide is curious, asks one question, and never grades. If something is wrong, it asks a question that leads back to the lesson's own explanation.
Why
Explaining something is one of the best ways to understand it, and teaching a character gives children a reason to explain.
What the research says
- Evidence: limited Students who taught a computer character spent more effort and learned more, but the study was with older children (Chase et al. 2009).
What we deliberately do not do
No coins for teaching, no grades, no voice stored, and the guide never becomes the source of facts.
Privacy from the start
Status: Live in the prototype. Consent switches, export and delete, retention rules, and a check that no tracker is added.
What it is
A parent account controls everything, with a separate switch for each child for each AI feature. A switch that is off hides the feature, and the server enforces it. Every stored item is tagged, so everything about one child can be downloaded or deleted. Speech is turned into text, and the voice itself is not kept.
Why
Photos of children, face measurements and voice recordings are some of the most sensitive data there is. The simplest protection is not to collect them.
What the research says
- Evidence: not established This method rests on law and regulator guidance rather than studies: the US COPPA rule as amended in 2025, the UK age appropriate design code, Israel's Privacy Protection Law Amendment 13, and the EU AI Act. This is our reading, not legal advice.
What we deliberately do not do
No photos of children to build a character, no face templates, no stored voice, no use of children's data to train AI, and no third-party trackers.
Telling a grown-up, and a time limit that ends gently
Status: Built, not yet released
What it is
Every chat with the guide shows that parents can read the conversations, with a "tell a grown-up" button. A parent can set a daily time and quiet hours. No countdown on screen, and the reminder comes only at a natural stopping point, never in the middle of a question.
Why
Children should know when adults can read, and have an easy way to reach an adult. Parents need limits that children experience as rest, not as a clock.
What the research says
- Evidence: not established Regulator guidance rather than a trial: the UK age appropriate design code asks for an obvious sign when monitoring is on and for wellbeing tools without nudging (ICO standards 11, 13, 15).
What we deliberately do not do
Tapping "tell a grown-up" creates no transcript or summary for the parent; the parent sees only that they asked, and when. There is no time limit by default.
Speaking to every child in the right words
Status: Built, not yet released. The first screens.
What it is
Hebrew has no neutral "you". Each profile is addressed with girls' words or boys' words. We write neutral wording first, and a whole sentence in each form where warmth needs it, never a slash.
Why
Children notice when the language is not speaking to them.
What the research says
- Evidence: not established Our own measurement, not a research claim: an automatic check confirms the girls' wording did not change and that the boys' wording contains no feminine address.
What we deliberately do not do
The address is a language setting only: it does not change items, themes, difficulty or anything else.
What we chose not to build
We test popular claims before we build on them. These are some we checked and rejected.
"Half right, half wrong is the ideal learning zone."
That is the target for tests. Practice systems for children aim higher and let children choose.
"Detect frustration from screen pressure or shaking and drop the difficulty instantly."
The cited studies used adults or sensors, not children in an app. In the EU, emotion recognition like this is prohibited in education institutions.
"Stop counting mistakes when the child is frustrated."
No study supports ignoring answers because of an inferred mood. Only rapid, effortless guesses are left out.
"Owning items keeps children from leaving."
Ownership is real value only if it is not an exit barrier. Children's creations never decay and leave with them in the parent's export.
"Let one sibling's progress unlock content for the other."
That creates obligation between siblings. Shared goals everyone can contribute to do the good part.
References are given as they appear in the research literature.