How the AI pipeline works

1. 1
   Parse

   The document is read. Scanned pages go through OCR and low-confidence pages are flagged rather than silently accepted.

2. 2
   Categorise

   The syllabus is classified and its strands identified — the thematic threads that run through it.

3. 3
   Deconstruct

   Learning outcomes are extracted and gates are proposed, each with prerequisites and a Bloom target. This produces the knowledge graph.

4. 4
   Schedule

   Gates are allocated across the available weeks, respecting prerequisite order.

5. 5
   Human review

   Everything above is presented for correction and approval. Nothing proceeds until a teacher approves — this is the gate in the pipeline, not a formality.

6. 6
   Author

   Per gate: lesson plans, four-stage Socratic scripts, question banks with misconception distractors, and study materials.

Extraction is cheap and fast; authoring is expensive and slow. Splitting them means the expensive half only runs against a structure a human has confirmed. Correcting a gate sequence before authoring costs a two-minute edit; correcting it afterwards means regenerating everything derived from it.

Beyond authoring, models do three things at assessment time: mark short and open-ended answers, grade scanned paper scripts against an extracted answer sheet, and grade long answer books against a rubric.

All three are advisory. A teacher’s override is final and recomputes mastery immediately.

• Clustering wrong answers into named misconceptions.
• Tagging questions to concepts, which is what the concept map is built from.
• Transcribing lesson recordings and voice feedback calls.
• Proposing teaching adjustments from analytics — each accept, edit or reject.
• Drafting report narratives.