WHO WE USE

Five companies, and what each one holds.

Nobody runs a website alone. These are the only outside companies involved, and this is the whole list rather than a sample of it.

Three that are how the site works at all

You cannot opt out of these and still have a working site, so we tell you about them rather than asking. That is the honest distinction: necessary things get disclosed, optional things get asked.

Supabase, the database

Holds everything: your email, your child’s first name and birthday, and every answer. This is the filing cabinet, and the rules about who can open which drawer are enforced inside it rather than by the website asking nicely.

Vercel, the hosting

Serves the pages to your device and sees the ordinary technical things any web host sees, like which page was requested and roughly where from. It does not hold your child’s answers.

Cloudflare, the domain name

Turns realadaptive.com into an address your device can reach. It does not see which page you asked for.

Resend, the email

Sends your six digit sign-in code and account notices. It sees your email address. It never sees anything about your child. There is no marketing email, so this is the only reason we would ever email you.

One that does three jobs, and one of them is optional

Anthropic, the AI that writes questions, writes lessons, and checks reasoning

It does three things. Only the last needs your permission, and everything still works if you say no to it.

It writes your child’s questions. Rather than everyone getting the same worksheet, questions are written for where your child actually is, about things they are interested in. Every one is solved again by a separate AI before your child sees it, and anything the two disagree about is thrown away.

It writes the lesson when something goes wrong. On the parts of maths nobody here has written a lesson for yet, this writes the short explanation your child reads after a wrong answer. It is a separate request, and what it is told is below.

It checks written reasoning. When your child is old enough to write out how they worked something out, this reads what they wrote and answers one question: does this show real reasoning, or is it the question repeated back? This is the part we ask you about, and the part you can say no to.

What is sent, and these limits are the point:

To write the questions: your child’s age in years, the interests you gave us when you set them up, and a summary of how they have been getting on — which skills they have practised, how many they got right, roughly how long they take, and the names of the mistakes they have made. That summary is what makes the questions land at the right level instead of at an average, and it is also what decides which skill comes next, how hard it is, how many times they have to prove it, and whether they can skip a skill they have already shown they can do.
To check their reasoning: only the words your child typed. This is the part covered by the permission we ask you for.
Never their name, never their date of birth, never their school or where they live, and nothing attached that says which child it is.
Anthropic is based in the United States, so this is a transfer of information out of the Philippines. We are telling you because that is a thing you are entitled to know before agreeing, not because it is buried in a rule.
It is never used to train AI.
Where a person here has written the lesson, that person’s version wins. The AI only puts it into simpler words for your child’s reading age, and a second AI checks the rewrite still teaches the same thing — if it does not, your child sees the original.
Where nobody has written one yet, the AI writes it. That is most of the maths above fractions, and until now it meant your child got a wrong answer marked wrong and no explanation at all.
Writing that lesson is a separate request, and it is told almost nothing. The question itself, the answer we already have, one wrong answer to explain the arithmetic of, the name of the skill, and your child’s age in years so the words are not too hard. It is not told how your child has been getting on, the names of mistakes they have made, or what they answered. Writing the questions needs those. Writing the lesson does not, so it does not get them.
Two things are checked before they read it. A second AI solves the question on its own and has to reach the same answer. And the explanation may never say what your child was thinking — not because a computer could not produce such a sentence, but because it would be guessing, and we will not put a guess about your child’s mind in front of them as though it were a fact. If either check fails, your child is told the correct answer plainly rather than shown something we have not checked. A model can be confidently wrong in front of a child, so nothing it writes reaches your child unchecked.
It never grades spelling, grammar or how well something is written. It asks only whether the ideas are there. A child writing in their second language, or a child who finds writing hard, is not penalised for it.

Say no and everything still works. Your child taps their reasoning from a list instead of writing it out, so nothing they write is ever sent. That is what saying no protects: their own words. Their questions are still written for them, the same as for every other child, and so is the lesson when something goes wrong. We ask you once, when it first matters, and if you say no we never ask again. You can change your mind either way in your own settings, and that is the only way it changes.

One thing no company sees

The distress check on anything your child writes runs on our own infrastructure, not on anyone else’s. That is deliberate: saying no to the AI service must never switch off the part that keeps a child safe.