A general-purpose model can explain photosynthesis. What it cannot reliably do is explain photosynthesis the way a Maharashtra State Board Class 8 textbook does, using the diagram that book uses, in Marathi, at the reading level that student has reached.
The mismatch is specific, not vague
Three gaps show up repeatedly when a generic model is dropped into an Indian classroom:
- Curriculum drift. Answers are correct but out of scope, introducing terms the syllabus has not reached.
- Example mismatch. Word problems reference contexts unfamiliar to the student, adding a comprehension barrier to a maths question.
- Language thinness. Regional language output is often a translation of English reasoning rather than an explanation constructed in that language.
What building custom changed
Training against the actual corpus — 65 years of textbook, workbook and teacher-guide material — closes the scope problem directly. The model knows what Class 8 has covered because it has been taught the same sequence.
Accuracy in an exam context is not just “is this true”. It is “is this true, in scope, and expressed the way this student has been taught to read”.
Measuring it honestly
We evaluate against past board papers and teacher review rather than generic benchmarks. A model that scores well on international reasoning tests but drifts out of syllabus is not useful to the teacher standing in front of the class.