Lecture Studio

Lecturers teach. AI transforms. Students learn.

Every lecture taught becomes structured, searchable, accessible academic knowledge — a corrected academic text, structured notes, a taught audio lesson, revision material, and one more entry in a knowledge base the course's own AI answers out of.

It is not a transcription tool. The transformation is:

VOICE → KNOWLEDGE → NOTES → AUDIO → REVISION

---

The pipeline

LECTURE  (mp3 · m4a · wav · video · a pasted transcript)
   │
   ▼
TRANSCRIPTION
   │
   ▼
AI ACADEMIC PROCESSOR  ─── three passes, three artefacts
   │
   ├── Grammar & language  ──► Corrected academic text
   ├── Knowledge extraction ─► the COURSE KNOWLEDGE BASE
   └── Structure & format  ──► Structured notes
   │
   ▼
AI TEACHING SCRIPT   ──► 15-MINUTE AUDIO LESSON   (a long lecture becomes several)
   │
   └──────────────────► REVISION  (flashcards · quiz · MCQ · exam questions ·
                                   terminology · what you must remember · 5-minute)

Each stage is its own artefact with its own version history, its own state and its own review. Nothing is a hidden step inside something else, because a lecturer cannot correct what they cannot see.

The workflow is never Lecture → AI → Student

Lecture → AI → LECTURER REVIEW → Student

The lecturer reads what the model produced, corrects anything it got wrong, approves it, and publishes it. Students then see Published by Dr —, and the university has a named academic answerable for every word. The platform enforces this rather than recommending it: an unapproved artefact cannot be published, and src/lib/domain/ownership.test.mjs is where that is held shut.

A correction makes everything built on it stale and offers to regenerate:

Recording → Transcript v1 → AI processing v1 → lecturer corrections → Published v2
                                                     └──► notes, script, audio,
                                                          quiz and flashcards
                                                          regenerate from the
                                                          corrected authoritative
                                                          version

The AI's job is transformation, not adjudication

The processing model works under a contract (src/lib/ai/contract.ts), prefixed verbatim to every transformation, and prompts.test.mjs fails if any stage is missing it:

ROLE: LECTURE TRANSFORMATION ENGINE. You are NOT a fact checker. You are NOT an academic reviewer. You are NOT permitted to correct factual claims, add information that was not in the lecture, or remove a claim because you believe it incorrect, controversial, biased, incomplete or unconventional. When uncertain whether a change would alter meaning, preserve the original wording.

It is not asked to correct anything. It is asked to perform a constrained linguistic transformation with an enumerated list of permitted operations — grammar, spelling, punctuation, sentence structure, filler, repetition, paragraphing, headings, bullets — and an equally explicit list of forbidden ones, factual correction and adding context among them.

Then a verification pass asks the only question worth asking about the result: did it introduce, remove or alter any substantive claim? Not whether the lecturer is right. The review screen shows the diff of claims.

If the lecturer taught that the Roman Empire fell because Christianity weakened its military, that is what the student receives. Whether it is sound is the lecturer's business, on the review screen, in their own words.

There is a constitution, and it is enforced in code. Seven articles — source preservation, meaning preservation, terminology preservation, no unrequested knowledge injection, no silent normalisation, traceability, explicit separation — each naming the files that make it true, with a test that opens them (src/lib/ai/constitution.ts).

The lecturer's terminology is authoritative. Yahuah stays Yahuah, never Jehovah, Yahweh or Lord; Yahusha HaMashiach is never Jesus Christ. The rule is in every prompt — and then the output is checked by counting, with no model consulted, because a rule a model is asked to follow is a rule nobody has watched. A substituted, dropped or respelled term is shown to the lecturer above everything else on the review screen, since the student who reads the substitution cannot see it.

The AI has permissions. Think of it as RBAC for AI: every model call is made as a role — transformation, verifier, course tutor, general explainer — and callAs refuses the call when the role's conditions are not met. A transformation prompt missing the contract will not run; the tutor will not run with an empty corpus; general knowledge is unreachable unless a student explicitly asked for it. Not "please behave yourself", but a defined permission boundary.

The model is not chosen yet. bench/ is a preservation benchmark — sixteen passages written to tempt a model into being helpful, scored mechanically. Claude is wired first; another vendor is one file against the same interface. The winner is the model that preserves best, not the one that knows most.

Who owns what

	Owns
Lecturer	The academic source material, and everything the AI makes from it. It is theirs; it leaves with them.
University	The course environment: faculties, departments, courses, who teaches, who is enrolled.
AI	Nothing. It transforms, and proposes. It never approves and never publishes.
Student	Consumes what was published on a course they are on, and asks the Course AI.

An administrator cannot edit a lecture, approve one, publish one, delete a recording, or read an unapproved draft. That is not a setting.

One lecture. Every language.

Teach once. Learn everywhere. A lecturer teaches in their own language; a student in Brazil, France, Kenya or China gets the same course — notes to read and copy, a lesson to listen to, a quiz, and the Course AI — in theirs.

MASTER — English, lecturer approved
   ├── Français   ├── Español    ├── Português
   ├── العربية    ├── 中文        ├── Kiswahili
   ├── Deutsch    └── Norsk …

Translation happens only after the lecturer has approved the original. Translating a draft multiplies one mistake into six languages and then asks a lecturer who reads one of them to find it. The gate refuses it.

Every language is a derivative of the master, never of another translation, so six languages are six renderings of one lecture rather than a chain that has drifted. The original governs: where a translation differs, the original is what was taught and what the student is examined on, and it is one click away on every translated page.

• Terminology crosses untouched. The same protection layer runs here, and for a sharper reason: a model translating Yahusha HaMashiach will otherwise reach for the target language's conventional name. It never sees the term — only a marker — and a translation that lost one is rejected.
• Checked without reading the language. Nobody here reads all nine, so the validator checks what is true in every language: the lecturer's protected terms, every figure the lecture states (١٩١٤ counts as 1914), the same headings, the same number of questions, and a length that is not absurd.
• Who read it is on the page. A lecturer cannot approve a language they do not read — the capability is held by a translation reviewer — so a translation nobody has checked says exactly that, every time it is shown.
• A correction to the original makes every translation stale, and says so rather than continuing to look approved.
• Audio in each language. "15-minute lesson" means about fifteen minutes: the same lecture is 15:00 in English, 15:12 in French, 15:20 in Arabic. Trimming the Arabic would mean cutting a sentence the lecturer said.

Quizzes are written once

Approved lecture → MASTER QUIZ → translation → localised quiz

Never translated notes → a quiz written from them: that gives the French cohort different questions from the English one, and they sit the same examination. Course content is one academic source; the student's language is a presentation layer over it — so the Course AI retrieves from the master and answers in whichever language the student reads.

The student's learning profile

MY LEARNING PROFILE
  Working language   🇫🇷 Français     — locked; the registry changes it, with a reason
  Voice              🎙️ Lecturer's    — yours, change it whenever you like
  Speed              1.0×

Two layers, deliberately unlike each other. The working language is the academic environment — notes, transcript, audio, quizzes, flashcards, the Course AI — set once, with no switcher inside a course. The voice changes only how it sounds.

And a lecturer's voice is never synthesised without their own authorisation: the capability is theirs alone, the default scope is translated audio only, a withdrawal is honoured at listening time, and where it is not enabled the option stays on screen saying so rather than quietly disappearing.

If twenty thousand students work in French, the course has one French version — not twenty thousand translations. Two students in one seminar must not be revising from different papers.

The Course AI

Restricted to one course's published lectures. Ask it:

• "Explain photosynthesis based on our lectures." — answered from the course, with the lectures cited
• "Which lecture introduced this concept?" — Lecture 06, straight from the knowledge base, with no model call
• "Give me a simple explanation." / "Now the university-level explanation." — same material, the register the student asked for
• "Create a 10-question test." — built from the lectures, not from the internet
• "Create a 15-minute audio revision covering lectures 1–6."
• …and it answers in the student's language while quoting the lecturer's own sentence beside its rendering, because that sentence is the one they are examined on.

And when the course does not cover it, it says so and names what the course does cover. A confident answer from outside the syllabus is how a student revises the wrong thing.

Asked from inside a lecture (Lecture Companion), the ladder runs this lecture → the course, and an answer found elsewhere says which lecture it came from.

Running it

npm install
npm run dev          # http://localhost:3000
npm test             # seven suites, no network, no keys
npm run build

It runs with nothing configured: the offline processor stands in for the language model, and stamps every artefact it produces so nobody approves machine prose believing a machine wrote it. Transcription and speech have no default vendor and say so instead of failing obscurely.

ANTHROPIC_API_KEY=...        # the real pipeline
ACADEMIC_AI_MODEL=claude-opus-5
ACADEMIC_DATA_DIR=./.data    # persist the demonstration across restarts
NEXT_PUBLIC_BASE_PATH=/academic   # when mounted inside another site

One sentence governs the repository:

The lecturer-approved master is the source of truth for what was taught. AI may transform its language, structure, translation and delivery, but it may not alter its substance.

• docs/ARCHITECTURE.md — the four products inside this one, the constitution that governs every AI operation, and where a sentence came from.
• docs/DELIVERY.md — how a lecture reaches a student, from a recorded multilingual lecture to live translated audio and lip-synchronised video. The last two are not built, and that page says so on every diagram.
• docs/BUILT.md — the audit of what is built, every figure recounted from the repository by npm run check:built.
• docs/GAPS.md — the other half of the truth: what is built, what is stubbed, what is a deliberate deviation, and what would have to be true to run a pilot.
• docs/INTEGRATION.md — mounting it inside an existing university system.