Solstra · a concept prototype of the AI-native classroom

The classroom that grows with every child.

The delivery layer a school runs on — units, live teaching, marking, assistants, a grounded notebook — rebuilt so AI proposes at speed, and a named human remains the marker of record, signing every mark that's recorded.

Enter the classroom → Read the thesis

runs gated on real Claude · wholly synthetic data · independent — not affiliated with Anthropic or Google

AI proposes. A human decides.

One rule carries the whole design: the model drafts, clusters and flags at speed — it never awards a grade, and accountability never passes from a person to a machine. No matter how good the model gets.

Why it exists

The AI tools sit beside the classroom. Solstra is the classroom.

Google Classroom owns the layer a school actually runs on. The AI tools that arrived recently bolt on from the outside — and the trust questions follow them in. Rebuild the layer itself, AI-native, and the guardrails become part of the architecture.

The current pattern

A side-tool, bolted on

Work is pasted out to a tool that sits outside the governed system — so data governance, safeguarding and accountability have to be re-argued at every seam.

The Solstra pattern

The layer, rebuilt AI-native

AI drafts inside the platform; the human-accountability core is load-bearing structure. Trust controls aren't a policy PDF — they're how the building stands up.

Six moments, working

Not a pitch deck. A working prototype.

Every moment below is live in the classroom — real interface, wholly synthetic pupils, one click away. AI in gold. Human decisions in petrol.

01 · LIVE SESSION

The class hits a checkpoint. The model proposes the patterns.

Wrong answers are clustered into proposed misconception patterns — never marked, never ranked. You review a pattern, approve a 60-second re-teach for just those pupils, and decide when the class moves on.

AI proposes · you decide

Checkpoint 3 · Selection — 24 synthetic pupils

0102030405060708 0910111213141516

"Selection = do BOTH paths" · 6 responses

Proposed pattern — awaiting your review · re-teach drafted for you to edit

AI proposed

Prog.

Support assigned to exactly those pupils — nothing is marked "correct". Progressing the class stays your call.

02 · MARKER CONSOLE

You mark first. Then the model shows its hand.

Human-first by construction: your judgement goes in before the AI's suggestion is revealed, and you cannot sign without your own mark. When the model reads a confidently-wrong answer as fine — you're the one who catches it.

Marker of record · human

Candidate 5A-04 · synthetic script

Your mark, against the rubric — first.☑ Names selection ☐ Explains the condition ☐ One path runs, not both

Your mark1 / 4entered before reveal

⭑ AI shadow suggestion — revealed after

Reads the answer as fine: proposes 4/4, "no rubric conflict flagged".

✗ Confidently wrong — the human catches it

✓ Override kept · Signed — marker of record

03 · ASSISTANTS

Helpers that give hints — never answers.

Teacher-built, subject-specific assistants with the guardrails baked in: hints not solutions, on-topic only, and "that's one for your teacher" when a question steps outside. Under-13 stays teacher-mediated, with no pupil accounts.

Guardrailed · runs on Claude

Code Breaker · Year 7 Python · teacher-provisioned

Pupil (synthetic)

what's the answer to question 3?

Code Breaker

Let's work it out together — look at the IF line. What does the condition check before either branch can run?

Pupil (synthetic)

who won the football last night?

Code Breaker

That's one for your teacher — I only help with our computing work.

04 · NOTEBOOK

Grounded, or silent.

The notebook answers only from the teacher's approved sources — and cites which one it used. When the answer isn't in the sources, it says so instead of guessing. No open web anywhere near children.

Cited · source-only

Grounded notebook · 3 approved sources

“What is selection?”

Selection means the program makes a choice: a condition is checked, and only one path runs.
◈ Source: Unit 3 · Selection (teacher-approved)

“Who invented the microchip?”

That isn't in your approved sources — so I won't guess. Ask your teacher, or add a source that covers it.

05 · STREAM

Nothing reaches the class until a teacher says so.

Teachers post; pupils ask and respond — and every pupil post lands as “awaiting review” until a human approves it. The same human-gated publishing discipline as a real audited classroom.

Human-gated

06 · TRUST LEDGER

Governance you can watch happening.

The rules aren't a policy document — they're controls, locked on, with a session log that records the AI's proposals and the teacher's decisions as they happen. Reviewable by anyone, at any moment.

Enforced, not optional

Trust Ledger · session log (synthetic)

🔒 Human marker of record🔒 Shadow mode🔒 Pseudonymised

14:02:11Model (simulated) — proposed 2 response patterns · pseudonymised

14:03:26Teacher — approved 60-sec re-teach · assigned to 6 candidate slots

14:07:44Teacher — overrode AI suggestion 4/4 → 1/4 · signed as marker of record

14:08:01System — “enable live AI marking” request blocked · DPIA gate

Responsible by design

Enforced in the architecture. Inspectable in the prototype.

The Legal Spine — the load-bearing rules the whole build protects. Written as a specification, because that's what they are; every row is a control you can watch working in the prototype's Trust Ledger and Notebook.

Marker of recordA named human, always. The model never awards a grade.

Shadow modeLocked on. AI runs alongside human marking — never on a live grade.

InputsPseudonymised. Names never reach the model.

Live useDPIA-gated. Blocked until the data-protection paperwork is complete.

Under-13Teacher-mediated. No pupil AI accounts, ever.

AnswersGrounded, or silent. Approved sources with citations — no open-web guessing.

◈ Watch it enforce — the Trust Ledger is live in the prototype

The human is the Sun — in charge of the machine. Every child, a star.

Sol + astra. Every learner is a star in their own right — different strengths, different needs, different ways of thinking — and the classroom grows with every one of them. It never ranks a child, and it never hands a child's future to a machine.

In practice: a named teacher signs every mark and approves every pupil post before the class sees it — and every assistant a learner can talk to is teacher-built, guardrailed, and teacher-mediated for under-13s.

Under the hood

It runs on real Claude — check for yourself.

The live demo calls a real Claude model through a gated, server-side proxy — the API key never touches the browser, a passcode gates the spend, and the marking engine falls back to a simulated model the moment anything is unavailable. Every response is labelled for exactly what it is.

The terminal isn't a mock-up: it queries /api/health on this domain as you read this — and you can run the same call yourself. Health reports the gate's configuration; the Marker Console in the prototype makes the model call itself.

“Claude” names the model the demo calls — nothing more. Solstra is an independent concept prototype by one teacher, and is not affiliated with, endorsed by, or connected to Anthropic or Google.

$ curl https://solstraclassroom.com/api/health
{
  "ok": true,
  "live": true,
  "gated": true,
  "model": "claude-haiku-4-5-20251001"
}
$

expected response — scroll here and this card queries the endpoint live

Who built it

Built by the person who runs this layer every day.

Saqub Hussain is a computing teacher who operates the classroom-delivery layer daily — an audited, human-gated workflow with AI assistants inside a governed workspace — and the author of a complete, exam-mapped Reception→Year 11 computing curriculum, published openly at plantbotcomputing.co.uk. Solstra runs a real Year 5 unit of that curriculum end to end.

The delivery layer isn't a theory here. It's a daily practice — rebuilt the way it should be.

“I run this layer every day, and I've drawn the line for where AI can and can't touch assessment — around children, under regulation. Solstra is what the AI-native version looks like when a human stays accountable.”

— Saqub Hussain

See it working

The whole thing is one click away.

Run the checkpoint. Catch the confidently-wrong mark. Ask the notebook something it has to refuse. Watch the ledger record all of it. Every claim on this page is inspectable in the prototype.

Enter the classroom → Read the thesis

concept prototype · simulated or gated-live AI · wholly synthetic, pseudonymised data