Vernacular-First Design for State Services

The English error string said 'ineligible — income proof stale'. The clerk, who was helping an older woman at a panchayat desk, translated it as a scolding. The woman left. The file closed. The model had been accurate in a language nobody in the room used for hard news. That is not a localisation bug. That is an exclusion error with grammar.

Vernacular-first means the workflow is born in the language of the form and the counter. English, if it exists, is a second skin for secretariat users. A toggle that translates a finished English sentence at the last mile will always sound like a toggle. Residents can hear it. So can courts, when the speaking order and the chat disagree.

This guide is dated 17 August 2026. It is not linguistic advice and not legal advice. Official-language rules are state law. Read yours. The Eighth Schedule is a constitutional list, not a backlog.

Design in the register you will be judged in

Pick the language of the speaking order first. If the officer must sign in Kannada, the draft must be born in Kannada. A translated English draft will carry English legal habits into a Kannada file. Officers will rewrite every sentence. You will call that 'low acceptance' and blame the model. You should blame the birth language.

Pick the language of refusal second. Refusal is where trust dies. 'I do not have that circular' must be sayable in the local register without sliding into English for the word 'circular'. If the model code-switches only when it is unsure, residents learn that English means bad news. That is a cruel interface.

Pick the language of the helper third. CSC operators and younger relatives will be in the loop. Design so a helper can operate without becoming the legal actor, and without the UI assuming literacy in the official standard. GIGW-sized type, spoken fallback, and official identity still apply.

• Templates before free text. A good Kannada deficiency memo beats a poetic one.
• Names and places stay in the script they were filed in. Do not 'normalise' a citizen's name through an English pipe.
• Numerals and dates follow the form, not the model's home locale.
• No public translation API on live personal data.
• Eval sets built with local clerks, not only with metro annotators.

Vernacular-first versus English-first with a toggle.

Choice	Vernacular-first	Toggle theatre
Draft birth	Language of the speaking order	English draft, machine-translated at send
Refusal	Local register, no English scare-words	English legal nouns in a local sentence
Eval	District clerks score fifty real memos	BLEU on a news corpus
PII	On-prem or SDC translation	Public translator 'just for language'
Script	Preserve filed script	Transliterate to Latin for the vector store

Models are not the language policy

A model that 'supports Indic languages' is a capability claim. Your policy is which language is authoritative, which is a courtesy, and which is forbidden for legal acts. Write that policy before you pick weights. Open-weight Indic checkpoints can still phone home, still leak, and still be bad at your district's register. Score them on your memos, offline.

Bhashini and similar public infrastructure may help on permitted paths. They do not replace a purpose tag. They do not make a health note safe to export. They do not settle whether a tribal language of the block must appear on day one. That last question is political and good. Answer it in the G.O., not in a model card.

Two counters after the same language G.O.

Objections you will hear — and what to do with them

We will add languages in phase two

Phase two is where rural users live. If they are not in phase one, you have built a secretariat tool. Say that, or change the phase.

One national model should handle all Eighth Schedule languages

It will not handle your forms equally. Score per language, per district register, per template. A weak language must refuse, not improvise.

Transliteration is good enough for search

Search maybe. Legal names no. Do not let a Latin pipe become the identity of record.

Spoken interfaces solve literacy

They can help, and they create an audio record that may be personal data. Design retention. Do not pretend a speaker is a policy.

A four-week playbook for one service and one language

1. Week 1: collect the form, the speaking-order style, and ten refusal lines a clerk already uses. That is the corpus.
2. Week 2: build templates in that language. Ban English nouns in refusal.
3. Week 3: score fifty drafts with two clerks who are not fans of the project. Pay them officially if needed.
4. Week 4: packet-capture to prove no public translator saw live packets. Then, and only then, discuss a second language.

File note you can paste

Subject: Language of agentic drafts and citizen interfaces — vernacular-first.

Drafts that officers will sign shall be born in the language of the relevant form and speaking order. Citizen-facing refusals shall use the local register and shall not drop into English for hard news. Live personal data shall not be sent to public translation APIs. Language capability claims will be scored on departmental templates, not on generic news corpora. GIGW 3.0 applies to surfaces residents see. This note does not amend the state's official-language Act.

This note is an internal aid. It is not legal advice.

Prcept AI will not ship you an English clerk with a toggle and call it Bharat. If we cannot refuse in the language of the desk, we will keep the agent silent. Silence is better than a cruel sentence.

Eval with clerks, not with BLEU

A generic Indic benchmark will not tell you whether your widow-pension refusal is an insult. Sit two clerks who were not in the project workshop. Give them fifty drafts. Ask three questions. Would you send this? What word is wrong? What would you write instead? Pay them through the official path. Keep their mark-up as the gold. If they rewrite the refusal every time, your template is wrong, not their taste.

Include mixed-script names and local place names in the fifty. Models that 'support Tamil' still mangle a village. That mangle in a speaking order is a legal defect. Score it as a defect, not as style.

When you add a second language, repeat the fifty. Do not assume the Kannada clerk's score travels to Kodava or to a coastal register. Language is not a checkbox on an Eighth Schedule poster. It is a sample you can fail.

This article is informational field guidance for Indian public institutions, not legal, linguistic or procurement advice. Confirm against state official-language rules, GIGW 3.0, DPDP, Bhashini terms if used, and counsel before you file it.

How to sequence this in a state, not a slide

“Vernacular-First Design for State Services” is a department problem. A P1 CIO/CTO should name the legacy system, the officer who owns the file, and the citizen charter clock before buying “vernacular government services design”.

A translation toggle on an English agent is not vernacular-first. Design the workflow in the language of the form, the clerk and the village. If the model cannot refuse in that register, it must not speak. Do not invent league tables of states. Read tenders and policies. Election Model Code of Conduct can freeze a rollout. NIC is a partner, not a villain. SDC readiness is GPU, power, ops and egress — not a logo.

• Audit the legacy store first.
• Keep mutation and money as officer actions.
• Map SLAs to the citizen charter.
• Budget change requests after go-live.

Close this loop before the next CAB

Put “Vernacular-First Design for State Services” on the next change-advisory or bid-opening agenda as a single line item with an owner. If it cannot earn a line item, it will not earn a control. The owner should be a P1 CIO/CTO, not “the vendor.”

Revisit the item when the model, the GeM term, the region, or the SI changes. “vernacular government services design” is not a one-time workshop. It is a watch item. Date the last check. Unsigned watch items are souvenirs.

What the next file must contain

“Vernacular-First Design for State Services” earns a line in the noting only if a P1 CIO/CTO can attach proof of “vernacular government services design.” A heading is not proof. A vendor slide is not proof. A workshop photograph is not proof.

Write three dated sentences: what was decided, who owns it after the next posting order, and when it will be re-checked. If you cannot write the three sentences, you are not ready to buy, to sell, or to go live.

Leave unsourced percentages out of the note. DPDP is not a blanket localisation statute. The November 2025 AI governance text is guidance, not an Act. CERT-In’s 28 April 2022 directions still set specified incident and log clocks. A PAC, when lawful, lives in GFR Rule 166.

• Name the designation that owns “vernacular government services design.”
• Attach one artefact a stranger can open next year.
• Record the instrument you are actually using.
• Revisit when the model, the SI, the notice or the posting changes.