What Nervur gives us, what's missing, the order of the work, and everything a new session needs to carry on.
Read in this order: this plan, then kartora-blueprint.html for the architecture, then kartora-alignment.html for the idea and the model. Where they differ, the Blueprint wins on architecture.
Where we stopped: the idea, the model and the architecture are aligned. Nothing is registered or built. We were about to start preparing the prototype.
Next steps: decide the open questions below, then write the Nervur statements, chunk the first math slice with its points and checks, and write the first five templates.
How this workspace runs: a session in repos/ is the orchestrator. Authoring happens in a session opened in the repo itself, commits go through the repo lobe, and Nervur's changes are handed to its authoring agent as one statement each: the fault, its source, the fix and its tests. kartora-docs/ is a plain folder, not versioned, and should move into the kartora repo once it exists.
Nervur has almost everything a prototype needs. Nothing blocks starting with the family organism on a small server and the child in a browser. The bulk of the work is Kartora's own code.
The two biggest unknowns aren't missing features. They're whether a browser home on a real site feels good to a 10-year-old, and whether the design work is bigger than it looks. Test those first.
| Need | Status | Notes |
|---|---|---|
| Family organism on a server or desktop | Built | Node installer |
| Child's organism in a browser | Built | @nervur-org/installer-browser |
| Invitations, links, standings, one-time references | Built | QUO-6, CASE-2. QUO-6 also lets a family ask for an invitation for another organism of its own. |
| Parent's AI by MCP | Built | nervur mcp reaches far organisms too. It runs on a desktop or server. |
| Lobes, cells, letters, alarms, jobs | Built | Core |
| Faculties, repo lobes, gates, releases | Built | |
| Pages drawn from a far lobe, a website's app view | Built | The window, LOOK-5 |
| Designs and websites at the edge | Built | quo.systems has a website |
| Calling a hosted model | Pattern works | A skin of network answers {job} and moves by scheduler rings, within the 10-second longest timeout. openrouter shows it, but it's private, so Kartora writes its own. |
| Family organism on an edge | Prove it | Cargo built and tested, no numbers measured (HOST-28). 48 MiB engine, 5-second budget. |
| Browser home inside a real site | Prove it | Built with tests. I found no site in this workspace embedding it yet. |
| Hosted model running on an edge | Prove it | Pattern proven on Node only |
| Edge pilot | Missing | Faculty and Node half unbuilt (HOST-7, HOST-26). Use a server, or lay an edge by hand. |
| File from the screen | Missing, no rule | Inputs in genome/package/src/standard/tree/allow.ts hold no file type. For now the parent's AI reads the PDF and sends the request. |
| Taking new versions by a rule | To confirm | LIB-8 mentions it. Move versions by hand at first. |
| Games as organs in a browser | Proposal | HOST-24. Templates are screen data drawn from the player's blocks. A lobe sends the screen data, never code (SCREEN-2). |
| Push, phone app, remote MCP, public log of browser code | Proposals | HOST-4, HOST-20, HOST-21, HOST-24, HOST-25, LINE-4, LINE-5, WEB-7, WEB-8. None blocks the prototype. |
The first mock-ups are in kartora-mockups.html. They show a Romanian family living in Germany. Ana, 9, follows the German school system and keeps Romanian as a heritage language; Matei, 12, follows the Romanian one. They share one subscription, AI and home. Each child has their own map across ten subjects, a shelf, knobs, an education country and an invitation. Every step of a session is a template with three parts, and Kartora's catalogue shows every chapter and template in public. Making them raised these problems the documents hadn't caught:
player lobe per child and a "Who's here?" screen, which is no lock. To decide whether a child needs a PIN.kartora, with its gate and harness. Everything Kartora develops lives there, and nothing depends on Bookarest Digital's private repos.family, child, player, author, graph, model, outbox, view.graph, check, signup, studies.frame, organism, lobe, data, app, loader), plus the coloured map and the shelf.| Step | What | Done when | Lead |
|---|---|---|---|
| 0. Prepare | Settle the decisions below. Start company, legal and trademark work. Find families and an advisor. | Name cleared, first families named, advisor found | All |
| 1. Paper test | Two weeks, 5–10 children, a small paper graph, a parent running sessions through a model. | We know if the method engages a child | Marcus, Razvan |
| 2. Two proofs | A browser home at a test site reaching a lobe on a server. One real template, filled with a child's interests, in front of a 10-year-old. Size the design work. | A child used it, and the design is sized | Razvan |
| 3. Graph | Chunk the slice, answering the open chapter questions on real content. | 15–20 chapters with relations, reviewed by the advisor | Razvan, advisor |
| 4. Thin slice | One child, one session, end to end: the parent's AI writes it by MCP, it lands on the shelf, the child plays it, results reach the family, a colour changes. | The loop runs once for real | Razvan |
| 5. Whole loop | Hosted author, fading, placement, frontier, approval knob, Kartora's relations and scoring, the outbox. | A session is proposed and scored with no parent acting | Razvan |
| 6. First families | 3–4 families set up by hand, then 5–10 children. Watch all three things we want to learn. | Several weeks of real use | Marcus, Razvan |
| 7. Towards real families | Move family organisms to edges and measure them. Carry the Nervur gaps. Payments and signup. | A parent sets it up alone | Razvan, Michael |
Not yet written as statements for Nervur's authoring agent. In order of need:
| Decision | Who | My suggestion |
|---|---|---|
| Math alone first, or with the mother tongue? | Razvan, advisor | Math alone |
| Who are the first children, how many, which language? | Marcus | 5–10 from your own circle |
| Is sharing metadata optional, or a condition of use? | All, lawyer | On for the first test, by agreement |
| Do results go to the family after each session? | Razvan | Yes, by a letter |
| Who pays, how much, for exactly what? | Michael, Marcus | A family subscription for the service |
| Do you need OAuth, or are invitations enough? | Razvan | Parents sign in at kartora.org. Children only use invitations. |
| Is the family organism always an edge, or an edge or a server? | Razvan | Either. Server for the first test. |
| Who owns graph content? | All | The advisor or a first hire |
| Where is the company registered? | Michael | Estonia is likely, not checked |
Graph questions still open are listed in the Alignment draft, under Open questions.
| Founder | Owns | First tasks |
|---|---|---|
| Razvan | Architecture, Nervur gaps, data model, faculties, product | Nervur statements, first math slice, the two proofs |
| Michael | Company, money, legal, investors, pricing | Country comparison, lawyer, trademark, domains, cost model, provider terms |
| Marcus | Families, brand, story, partners, teachers | 5–10 children, advisor, parent interviews, the story, a holding page |
Two risks: Razvan is the bottleneck across Nervur and Kartora, and nobody owns graph content yet.
Kartora has its own organisms and kartora.org, and knows its customer, the parent. The family's organism is the stable ground, where the children are declared and their data is kept. A child's browser organism is a stranger until it receives the family's invitation. Example:
Kartora doesn't keep children's names, emails, addresses or other details it doesn't need. It sees what it needs to run the service, such as which chapters are asked for and how checks go. It never promises complete privacy. Metadata can sometimes identify someone, and that's accepted.
What Nervur brings is boundaries held by structure, and relations the family controls. SaaS can connect the same people, but there the boundaries are a vendor's promise.
The model never sets the colour. Fixed check questions from Kartora, generated from patterns where possible, are scored by plain code, so colours mean the same in every family whichever model it uses. Free answers are a secondary signal. Check answers stay at Kartora.
.eu, .ai and .org free. WHOIS confirmed kartora.org unregistered on 7 November. Nothing is registered..org and .com unregistered and no collision found.cartora.org was in redemption, and cartora.com is taken.These come from memory or secondary sources. Don't rely on them yet.
claude -p in a family organism. Check OpenAI and the others too.