Kartora

Why, vision, mission

First draft Proposed

Why

Every child can understand, but school gives them all the same path at the same pace and rarely checks that they did.

Vision

Every child grows up holding a living map of what they understand, theirs for life.

Mission

Give every child a patient tutor that meets them where they are, never lies, and answers to their family.

Line

Learning is an infinite game.

The problem

School has kept one shape for 100 years: manual, narration, homework, test, then the same again next year with more on top. Knowledge flows one way and the child has little say. Technology changed the medium, chalk to PowerPoint, and left the method alone. Many teachers are weak, and phones now compete for every child's attention.

School also won't change quickly, and children still have to pass it. So Kartora starts beside school and helps with it, then takes on more over time.

The model in five ideas Agreed

An invented example of one child's map. Faded means new, full colour means close to 100% assimilated. The dashed link crosses into Geography.

The graph

The graph is the core of Kartora. It is a pure knowledge base kept by Kartora's supervisors. It says what is true and how things relate, and nothing about how to teach them, who is learning, or which school or country they are in. Agreed

• It is fixed within a release and changes by new versions, as software does. Agreed
• Chapters come from studying what schools teach everywhere, stripped of year and country. Agreed
• A chapter about one place, such as the history of Romania, is an ordinary node: true everywhere, needed by some. Proposed
• A language like German is its own discipline, taken only by some children. Proposed
• The product is built for every discipline from the start. Math is only where the first content goes. Agreed
• Each chapter lists its points, the things it contains, each with a knowledge shape: a fact, a concept, a rule, a procedure, a comparison, a classification, a sequence, a cause. Templates say which shapes they can carry, so choosing one never needs the model to judge the knowledge. Proposed
• Each chapter also lists its known misconceptions and how it's measured: its check questions, scored by code. Proposed
• The graph is public, as Kartora's catalogue. Anyone, a parent or a teacher, can check what each subject and chapter contains, its points, misconceptions and relations, seen through any country's lens, and see every template with examples. Check questions stay private, so they keep their meaning. Agreed

A good chapter Proposed

• Has a clear beginning and end, with no mention of age or country
• Makes sense with one colour on it
• Can be evaluated
• Can grow in content without being cut again

Disputed content Agreed

Where facts are truly disputed, a node holds variants, each saying whose position it is. A default can come from the country, and parents or teachers can adjust it. The session is honest that the point is disputed.

The lens Agreed

• Age is a reference, not a limit. It says what is typical. A curious 10-year-old can go to limits, and the bar rises for that child.
• The country is the lens, and the language comes from it. Knowledge is the same everywhere; the education country decides how it's presented. It sets the language of lessons, which is also the mother tongue subject, which languages are foreign, the typical ages and what comes first. In Romania, English and French are foreign languages; set France, and French becomes the mother tongue.
• The lens is more than a language. It also sets notation and units: 2/5 and 0,4 in Germany and Romania, 0.4 in the UK; 12 : 4 against 12 ÷ 4; euro, lei or pounds; kilometres or miles. And the order: Ancient Rome is read with Dacia first in Romania and with the Limes first in Germany. Proposed
• One lens per school system and language. Belgium has one lens for French and one for Dutch. Where regions differ, as Germany's Länder do, the lens names the region for its typical ages. Proposed
• Family default, each child their own. A Bulgarian family in Germany can have one child at German school and one following Bulgarian education, and a child can switch. Switching keeps the map, because the graph doesn't change. A language kept at mother-tongue level outside the country, like Bulgarian for a child at a German school, is a heritage language.
• The parent's own screen language is a separate setting. We still have to confirm that nothing country-specific belongs in the graph.

Templates and sessions Agreed

Knowledge and method are two different things. The knowledge sits at Kartora, in the graph. The method is how it reaches one child: templates, filled by the AI with what that child likes. That's what keeps a child interested, so it's as fundamental as the graph.

A template is a kind of step. Some present a topic: a story, a comic, an explorable picture, a timeline. Some are exercises: choose, fill the word, build, sort, order, spot the mistake, teach back. Kartora's check measures. A quest takes it away from the screen. Kartora ships them, versioned with its releases, and each release adds more and better ones.

Every template has three parts, always:

Each template also carries its harness: the rules the AI follows when it fills it, and what plain code checks afterwards. Every claim comes from the chapter's points; a wrong answer is a listed misconception; numbers are computed by code; the check's words are never changed. So the AI prepares the session, the parent decides only what is learned and approves, and the template holds the rest.

The check is the same for every child. Only the line before it can belong to the story. That keeps colours comparable across children, families and countries.

A session starts in one of two ways: the AI proposes it from the child's map, or the child asks, for example with a worksheet from school. The material is always mapped onto the graph first. A typical session: a story to present, two or three exercises with branches for likely mistakes, Kartora's check, a teach back, a quest.

The child decides how, never what: their interests, colours and names shape every template, and they can start sessions. What is learned is set by the graph and the supervising adults. The child's part never leaves the family.

The profile

• Each child has a colour on each chapter, from 0 to 100. It's absolute: the same scale for every child, with age only used to read it. Agreed
• The child, parent and later the teacher each get a view of it. Agreed
• Sessions record evidence: what was asked, answered and touched, and with which method and AI model. The colour is calculated from that evidence, so a better evaluator can recalculate every child's map later. Proposed

People and who has a say

Knobs and hard rules Agreed

Behaviour is adjusted by knobs, each with an owner. The parent's knobs are the allowed bias every template reads: the education country, how deep to go, how much interests flavour a session, themes to avoid, which disputed view comes first, time in a day. New ideas can become new knobs without touching the data model.

The hard rule: Kartora never teaches a lie, whoever asks. Knobs change emphasis and coverage, and they stop at the truth.

Built to grow Agreed

An upgrade must not force a change of the data model. Each part changes alone and meets the others at a few stable points.

Stable points Proposed: chapter IDs and releases, one template contract with its three parts, evidence records, colour derived from evidence, and the AI model behind one seam.

Think of it as reinforcement learning for humans: an inner loop per child, and an outer loop where experts learn what works across children. Two warnings come with that: anything optimised against a metric learns to game it, and a child has only one childhood to experiment on.

On Nervur Agreed

Kartora is a use case of Nervur, and the two go in parallel, so it is designed as if Nervur is finished. Nervur handles the relations between children, parents, teachers and researchers, and the devices they use. The data model doesn't depend on any device. Children will probably use an app. A game can be an organ, and how organs reach a general app is Nervur's question.

It uses public Nervur alone. Every faculty and organ it needs is developed inside Kartora, with nothing from Bookarest Digital.

Two stages. First it's an emulated SaaS: people register an account at kartora.org, signing in by OAuth, and Kartora keeps those accounts on Nervur as a keeper. Later, as Nervur grows, it also offers real Nervur accounts.

Privacy is why Nervur was chosen. Families can hold their own organisms. Kartora's organisms hold the graph and scoring alone, with no children's names or accounts. The emulated SaaS is only a transition. Agreed

Each side keeps its own half. The family's organisms hold the child's name, preferences, uploads, session text, evidence and scores. Kartora's hold the graph. Proposed

Keep how someone signed in separate from everything else, so the second stage is a swap and not a rebuild. Proposed

Kartora may receive metadata, never names or details of a child, so researchers can work and the system can be evaluated. Teachers can receive their own kind of information. Faculties on the family's side filter what leaves, and evaluation may be done by Kartora on metadata. Agreed

How that could work Proposed: one outgoing filter on the family's side, with a separate relation and view for each receiver. Only fixed fields leave, such as chapter, check question, right or wrong, time band and method version. Free text never leaves. The parent can read exactly what left. Kartora keeps the check answers, scores the metadata and returns the colour.

What Nervur brings. SaaS can connect teachers, parents and researchers too. Nervur keeps the privacy and the boundaries of what each one may see, and puts control of the relations in the family's hands. In a SaaS those boundaries are a vendor's promise. Here they're the structure, and no party holds the whole child. Agreed

Bring your own AI Agreed

Sessions run in the family's organism, driven by the family's AI. The parent chooses the model and decides how much data that provider sees. Kartora doesn't own session data, and may offer its own trained models for some parts later.

The shape Agreed:

• The child has a browser organism, on a phone or desktop, controlled by the parent's organism through relations and standings.
• The parent works in their own ChatGPT or Claude subscription, as an adult, and reaches the organisms by MCP. From there they control sessions and knobs.
• The child never touches ChatGPT or Claude. They only ever work in their organism.
• The parent can read every conversation and log, and what happens on the child's organism stays there.
• Kartora also uses MCP, to pilot the graph's content.

Sessions are prepared first Agreed: the parent's AI authors each session ahead of time, with its explanations, questions and branches for likely mistakes, and writes it into the child's organism. The browser plays it with no model live, plain code scores what it can, and the AI reviews the results at its next run. The child can also open sessions, within what was pre-approved.

How that could work Proposed: a shelf of ready sessions the child picks from; requests and uploads that wait for the AI's next run; and an approval knob, from approving each session to automatic within opened chapters.

The family also needs models it controls that can run with no parent at the keyboard, such as OpenRouter or hosted models on the parent's account, or claude -p authorised in the family organism. Agreed

Connecting the dots Proposed: four organisms and one seam.

Parents have organisms locally, and use an edge as a keeper. They control the Cloudflare account from the local organism, with the edge pilot. Agreed

The seam is the author: anything that can turn a request into a session. It may be the parent's AI by MCP, a hosted model, claude -p or a local model, and the family's knobs choose which. The child's organism never knows which one wrote a session. A request flows from child to family, to an author, through approval, and back to the shelf.

Still to settle Open: how a child can be present at an organism the parent controls, how simple the edge setup can be for a non-technical parent, whether live sessions come later, and which models are good enough.

How the model maps onto Nervur hasn't been designed yet.

The MVP

There is no finish line and no "success": it's an infinite game of updating the profile against the graph. Agreed The first children should still teach us three things: Proposed

• A child comes back without being pushed
• Sessions and uploaded PDFs land on real chapters
• The colour matches what a parent sees and how the child does at school

Ideas on the table Proposed

All of these interested you. None is decided. Most of them become templates.

Method

• The child teaches a confused younger character: the teach back template
• Off-topic questions become paths on the graph
• Chapters list known misconceptions, so a wrong answer says what to fix
• The colour fades without practice, like a garden
• The child says how sure they are before answering

Motivation

• The map is the game: fog of war, far goals in view, and the child chooses the route
• World quests beyond the screen
• One sentence a day for parents in place of a dashboard they won't read

Risks to design for

• Homework cheating: the AI must never just solve it
• "Never lies" needs checks outside the model, such as plain code for math answers
• Attachment: a tutor that sends the child back to people, not a friend
• Interests as bribery: the child's world must carry the idea, not decorate it, and the method should widen interests, not only feed them
• Stereotypes: interests come from the child and the parent, never guessed from a name or a gender

What Nervur makes possible

• A map for life, with ownership passing from parent to child at 18
• Children teaching children, with parents approving each relation
• A narrow standing for a relative or expert on one chapter
• Families sharing evidence with researchers on purpose, by a relation they can end

Testing cheaply

Try the method by hand for two weeks with 5–10 children, a small paper graph, and a parent running sessions through a model.

Open questions Open

1. Math alone first, or math and the mother tongue?
2. Who are the first children, how many, and in which language?
3. What does "quality of relation" hold: type, strength or both?
4. If a chapter gains content, what happens to a child at 100%?
5. Does evidence on one chapter affect the chapters linked to it?
6. How much does one mistake count against each chapter it touches?
7. Where do concepts that cross chapters go, such as the Industrial Revolution? Does each discipline have its own way of cutting chapters?
8. Do history's chapters fit in 100–200 if the graph holds every country's history?
9. Does the graph hold finer points inside a chapter's content, so evaluation can name what was missed?
10. What happens to uploaded material that maps to nothing in the graph?
11. Is the graph built by hand, or drafted by AI and approved by people?
12. Who maintains the age reference data, and where does it live?
13. What exactly counts as disputed, and how is a disputed point evaluated?
14. Does a parent open whole disciplines, or single chapters too?
15. What rules apply to children's data and to sending it to model providers?
16. Who pays, and for what?
17. Who curates each country's lens: typical ages, emphasis, chapter names and check questions in its language?
18. Which countries come first?
19. Which knowledge shapes does the graph need, and does every point get exactly one?
20. Who invents templates, how is a new one tested before release, and which come first?
21. How are pictures made for the child's world: a library drawn by Kartora, or generated and checked?

The company

Founded by three friends, with Europe first and the world as the goal. The country of registration isn't decided, and Estonia is the likely choice. Agreed

Still to answer Open: who brings learning science and child psychology, and who pays for an account and how much.

Nervur is a tech stack for Kartora, used under its Apache-2.0 licence, with no other link between the companies. Agreed

The name

Kartora is chosen. Agreed It comes from map words such as Karte, carte and kartë. You found .eu, .ai and .org free, and WHOIS confirmed kartora.org unregistered on 7 November 2026. Nothing has been registered.

A formal trademark search of the EU, WIPO and UK registers is still needed. Open The first thing to clear is CARTORI, a UK application filed in April 2026 for software, which sounds very close.