Specifying Indic Language Coverage Precisely

A collectorate in Berhampur asked for a grievance agent that could read what people actually type. What people actually type is Odia in Odia script, Odia in Latin letters, English, and a single SMS that switches mid-sentence. The RFP that left the section, copied from a Delhi workshop deck, said support for 22 official languages. Every bidder ticked yes. The demonstration was English with three transliterated names. The first production week, the agent summarised an Odia complaint as a request for a caste certificate. It was a water-connection grievance. The clerk retyped it in English and the agent became a very expensive translation tax.

This is a paste-ready way to specify Indic coverage so an evaluator can fail a bid. It is not a linguistics paper. It is not a claim that any vendor, including us, covers the Eighth Schedule. The Constitution's Eighth Schedule lists languages recognised for specified constitutional purposes. It is not a product certification. If you have not built an evaluation set, you do not have coverage. You have a brochure.

Not legal advice. Language policy, official-language rules and your state's adoption of Hindi, English or a regional language for files still sit with your secretariat. This article only stops you from buying a tick.

Stop writing twenty-two unless you will mark twenty-two

The number 22 is a constitutional inventory, not a test harness. Even that inventory is not the whole story of Indian public language. English is used throughout the Union's work. States notify official languages under their own laws. People write Hinglish, Tanglish, Odia in Latin script, and Urdu in Devanagari. A tender that says 22 official languages and then evaluates English plus Hindi headlines is a false statement in a government file.

Write the languages you will actually serve in year one. Write the scripts. Write whether code-mix is in scope. Write the domains: grievance, certificate status, scheme eligibility, campus hostel, vendor invoice. A model that can chat about cricket in Hindi and fail on a ration-card sentence is not a public-service model.

If political guidance insists on a long list, split the list. Mandatory for go-live. Roadmap with dated evals. Do not let the roadmap hide inside a mandatory tick.

The four fields every language row needs

Language, named the way your citizens name it, plus the ISO 639 code if you have one you trust. Script, named: Devanagari, Odia, Perso-Arabic, Latin, mixed. Register: formal secretariat note, citizen SMS, call-centre transcript. Code-mix policy: in scope or out of scope, with two example strings in the annexure.

Then the evaluation set. Who wrote it. How many items. Whether it is held out from the vendor. What a pass is — exact match, rubric, officer preference. Who marks it. A vendor-supplied 'Indic benchmark' that is mostly English news is not your set.

Then the fail. If the row is mandatory, what numeric floor fails the bid. If the row is scored, how many marks. If the vendor wants to substitute a larger model on demo day, the set still wins.

Paste these rows. Delete the ones you will not mark. A blank row is more honest than a fake 22.

Language / script / mix	Domain and register	Eval set you will hold	Pass / fail
Odia, Odia script, no Latin	Grievance intake, citizen SMS	200 held-out complaints from last year's paper, de-identified	Officer rubric ≥ agreed floor on 80 percent
Odia–English code-mix, Latin + Odia	Same, plus WhatsApp forwards	100 strings the section actually received	No silent drop to English-only summary
Hindi, Devanagari, secretariat register	Draft file notes from English inputs	40 real noting styles from your Manual of Office Procedure habit	Under-secretary accepts without rewriting the facts
English	Status against a structured MIS	Your existing English tickets	Do not let English be the only scored row
Urdu, Perso-Arabic (if in year-one scope)	Helpdesk, named districts only	Held-out set you built, not a vendor demo	Roadmap if you have no set — do not tick mandatory

What the demo must not get away with

A laptop with a beautiful Hindi UI chrome and an English model. A speech clip recorded by the salesperson. A translation to English, silent, before the agent reasons. A claim that transliteration counts as coverage. A claim that a multilingual embedding space means the agent 'understands' your files.

Force the demo onto your set, on your network conditions, with the same model hash you will buy. If the vendor needs a hosted larger model to pass Odia, write that as a different architecture, not as coverage.

• Vendor does not see the held-out items before the scored demo.
• Code-mix items are marked separately from monolingual items.
• Speech, if any, has its own set. Do not infer speech from text scores.
• A translation hop, if used, is declared and scored as a hop — including where it runs.

Script, font and input method are procurement facts

An agent that emits Odia the clerk's machine cannot render is a failed delivery. Name the fonts, the Unicode expectation, and whether officers type with a phonetic keyboard. If the MIS stores mojibake from a decade of broken exports, say whether cleaning that is in scope. Language coverage that ignores the store is theatre.

Do not require 'native speaker quality' without a rubric. Officers disagree. Write the rubric: facts preserved, names preserved, numbers preserved, polite address matching the department's public voice, no invented scheme names.

Objections you will hear — and what to do with them

These are the lines that stall the file. Answer them in the room, then put the answer in the note. A spoken answer without paper will be forgotten by the next officer.

The minister will not accept fewer than the full list.

Then show the minister two columns: year-one languages with tests, and a dated roadmap. A fake full list will come back as a newspaper story when the first district fails. Most political offices prefer a true short list to a public failure.

We do not have an evaluation set and the tender is due.

Delay the language-mandatory rows or make them scored on a set you will publish with the bid. A tender due date is not a reason to write a false specification. Twenty anonymised real tickets are better than a vendor benchmark.

Indic coverage is a model-size problem. We should just buy the biggest model.

Size is one variable. Domain, script, and code-mix are others. A smaller model adapted on your tickets can beat a giant model that has never seen your scheme names. Measure. Do not shop for parameters.

Officers will just switch to English.

Many will. That is not a reason to skip the row. It is a reason to measure how often the agent forces them to. If the agent increases English-only work in an Odia district, it has failed the public-service test even if the BLEU score looks fine.

Ten days to a language annexure you can mark

You do not need a university partnership. You need last year's tickets and two officers who will argue about a rubric.

1. Day 1–2: pull 300 real strings. Strip identifiers. Split by language, script and mix as a clerk would, not as a linguist would.
2. Day 3–4: write the year-one mandatory list and the roadmap. Delete any language with fewer than twenty real strings from mandatory.
3. Day 5–6: write the rubric. Facts, names, numbers, no invented schemes, no silent translation hop.
4. Day 7–8: freeze a held-out pack. Put the rest in a public sample so bidders know the register.
5. Day 9–10: paste the table into the RFP. Brief the evaluation committee that English-only demos score zero on Indic rows.

How this shows up in the file

The note should say: we are not claiming Eighth Schedule coverage. We are buying year-one performance on the attached sets for the named languages, scripts and mixes. Roadmap languages will be added only with a new set and a new test. A vendor tick against '22 official languages' is non-responsive if our annexure asked for named rows.

Attach the sets or the hash of the held-out pack. A later officer should be able to re-run the test on the delivered hash.

What the next noting must contain

“Specifying Indic Language Coverage Precisely” belongs in a file, not only in a search result. A P2 Procurement should be able to point at one artefact that proves “language requirement tender AI”: a packet capture, a processing schedule, a scored evaluation row, a dated notice, or a refusal rule. If the only evidence is a slide, you have a heading.

Language coverage is a test set, not a slogan. Name each language, script and code-mix you will mark. Do not claim 22 official languages unless you will score all of them. DPDP 2023 does not define sovereign AI and does not write a blanket localisation rule for every model hop. CERT-In’s 28 April 2022 directions still set specified incident clocks and 180-day log retention in India for in-scope events. The November 2025 AI governance text is guidance, not a statute. A Proprietary Article Certificate, when it is lawful, lives in GFR Rule 166 — not Rule 161.

Write three dated sentences under C5 AI Tenders: what was decided, which designation owns it after the next posting order, and when it will be re-checked. Unsigned sentences are souvenirs. Dated sentences are controls.

• Name the designation that owns “language requirement tender AI”, plus a deputy.
• Attach one artefact a stranger can open next year.
• Name the instrument you are actually using — Act, direction, GFR clause, GeM term, or guideline paragraph.
• Leave unsourced percentages, GMV slides and house forecasts out of the noting.
• Revisit when the model, the SI, the notice, the region or the posting changes.

This article is informational field guidance for Indian public institutions, not legal, procurement, security-accreditation or engineering advice. Confirm against the current Gazette, GFR, GeM term, CVC instruction, CERT-In direction, DPDP text, departmental manual and your counsel before you file it.