In a Nepali school, Social Studies is taught in Nepali, printed in Nepali and examined in Nepali. The textbook is in Devanagari, the SEE paper is in Devanagari, and the answer the child has to write in the exam hall is in Devanagari. A practice worksheet for that subject written in English is not a matter of taste. It is the wrong artefact: it teaches vocabulary the marker will not be looking for, and it asks a student to do a translation exercise on top of the actual subject.
The same has always been true of the Nepali subject itself. A Nepali lesson explained in English is not a Nepali lesson.
The subject decides, not a setting
So we stopped treating the language as a preference and made it a property of the subject. When anything is generated — a worksheet, an exam paper, a lesson plan, a quiz round, a set of slides, a tutor reply — the server looks at which subject it is for and decides the language there. Social Studies and Nepali come back in Nepali. Everything else comes back in English unless someone asks otherwise.
Doing it this way means a school does not have to remember to configure anything, and a mis-set control on one screen cannot quietly produce an English Social Studies paper. It also means the rule survives a school's own naming: whether a subject is stored as Social Studies, सामाजिक अध्ययन, or a romanised spelling, it is recognised. We matched narrowly on purpose, so a neighbouring subject like Health, Population and Environment is not swept into Nepali along with it.
Nepali is not one register — it is three
This is the part that took the longest to get right, and it is why the rule could not simply be a Nepali on-or-off switch. A printed paper, a quiz round and a chat are not the same kind of writing, and they should not read the same.
- Fully in Devanagari — for anything that stands in for a printed document: a worksheet, an exam paper, a mock test, a lesson plan, a weekly plan. It is marked against the book, so it is written like the book, headings, options, answer key and explanations included.
- Mostly Devanagari, keeping established English terms — for a quiz round, a puzzle and the tutor's own replies. About nine words in ten are Nepali; the rest are the terms a Nepali student already meets in English, such as GDP, UNESCO or SAARC.
- Both languages, line by line — an option available on worksheets only, where a teacher wants an English line and the same line in Nepali on the page.
The middle one is the interesting choice. A game and a conversation are not marked, and a coined Nepali equivalent for a term the next chapter prints in English helps nobody — the student would have to unlearn it. So in a played round or a tutor exchange we keep the term the textbook will use, and put the reasoning around it in Nepali. In a printed paper we do not, because that paper is standing in for the real one.
A mock test is the clearest illustration. A quiz round and a mock test are built by the same machinery, but only one of them is pretending to be an exam paper. The mock test gets the full Devanagari treatment; the practice round gets the conversational mix.
A teacher may override it. A student may not.
Who is allowed to change the language is a separate question from what the default is, and the answer is different for the two audiences.
A teacher building a worksheet, an exam paper, a lesson plan or a set of slides can still ask for English Social Studies. Bilingual classes are real, and the teacher standing in front of the room is the one who knows whether this class needs it. So those screens show a language picker, and an explicit choice wins.
A student cannot. The quiz page used to carry a small English–Nepali toggle, and we removed it deliberately. It let a child opt out of the language their own exam is written in — and because it was a single saved preference applied to every subject, a student who flipped it once for Social Studies was quietly asking for Nepali Science quizzes too. Student-facing screens now show the language as a label rather than a control, because a control that the server is going to ignore is worse than no control at all.
Nobody, teacher or student, can switch the Nepali subject into English.
Three things we deliberately did not translate
A language rule applied without judgement breaks things that were working, so there are carve-outs, and each of them looks like an oversight until you see what it protects.
When a teacher photographs a page of handwritten homework, the transcription of that page reproduces exactly what the child wrote, in the script and language they wrote it in. It is evidence, not prose. A teacher who opened the photo to look at a mistake should be reading the mistake, not a tidy translation of it.
A Nepali answer is marked on its content. Answering in Nepali is never treated as an error, never lowers a grade and is never flagged as an issue for the teacher — regardless of which language the feedback around it is written in.
And a child who types 12 where the answer key says १२ is right. On a shared school tablet a Nepali keyboard is not a given, so the two are compared as the same number. The sheet is testing whether the student knows the answer, not which keyboard they had in their hands.
Holding the line against the material itself
One last practical detail. The material the platform works from alongside a question — a scanned page, a teacher's notes, an extract retrieved from the curriculum shelf — carries a language of its own, and a page of English can pull generated text back into English out of sheer momentum. So the language instruction is stated as overriding the language of anything quoted beside it, and it names the parts that otherwise drift: the headings, the option labels, the answer key, the explanation. A rule stated once loses to a page of counter-example.
None of this is a language feature, really. It is a curriculum one. The child sitting the SEE paper will answer in Nepali, and everything we put in front of them beforehand should be preparing them for that paper rather than a version of it we found easier to produce.