Multilingual communication generator

The generator's job is to produce a notice in the recipient's language and at their literacy level that a communications officer can defend. Quality here is not fluency alone. A fluent sentence that adds a payment figure, a new deadline, or a courtesy promise the English source never made is a failed draft.

Every claim in the output must cite the approved English source: the locked paragraph, field, or sentence it came from. If the source is silent on a point the template expected to fill, that field stays empty. Empty is correct. Guessing is not.

A human still sends. The model does not publish, queue a mail run, or mark the resident as notified.

This page is for the officer who owns the wording of service notifications: appointment letters, status updates, collection instructions, and similar outbound text that must match the authorised English version. It is not a substitute for automated eligibility determination. Amounts and decisions belong in the system of record. The generator only restates what that record already authorised in English.

Do not translate a working draft, a ticketing comment, or last week's template with a new date typed in the subject line. Lock a single approved English notice: version, date, and the programme that owns it. That locked file is the only source the model may read for facts.

If the English text is still being negotiated with legal or the service team, translation waits. Parallel drafting in two languages is how amounts and dates drift.

Once locked, freeze the fields the notice is allowed to contain: what happened, what the resident must do, dates that appear in the source, references, and the contact route the source names. Anything else is out of scope for this run.

Literacy adaptation is part of the same lock. If you need a plain-language or Easy Read version, decide that against the locked English, not against a second interpretation. The plain version must still cite the same source sentences. Shortening a paragraph is allowed. Softening a requirement into optional advice is not.

Run the generator against the locked source and the target language (and literacy level) only. Instruct it to attach a citation to each factual clause: a pointer back to the English sentence or field. Reviewers should be able to walk the draft without hunting.

Where a template slot has no source text, leave it blank. Do not backfill from memory of similar letters, from a previous campaign, or from what residents usually need to know.

Three failure modes show up in review more than any others.

Adding a promise that is not in the source. A Spanish draft that adds "we will collect any leftover items the following week" when the English notice only booked one slot is a new operational commitment. Delete it. If the service wants that promise, it goes into a new locked English notice first.

Inventing a deadline. If the English source does not state a respond-by date, the translation must not mint one to make the letter feel complete. A fabricated deadline creates appeals, missed rights, and mailbox traffic the programme did not plan for.

Treating the draft as sent. A completed translation sitting in a workspace is not notification. Until a named officer releases it through the live channel, the resident has not been told. Do not close the case, start a clock, or tell another team they have been informed on the strength of a generated file.

A bulky-waste team locks this English notice: collection is booked for Thursday 12 March; items must be at the kerb by 7am; the booking reference is BW-4419; if the crew cannot access the street they will not return automatically; questions go to the number printed on the notice.

The generator produces a Polish version and a plain-language English version. Each of those five facts carries a citation to the locked sentence it came from. The Polish draft does not add a refund if the crew misses the slot. The plain-language draft does not add "please reply by Friday" to prompt action. The contact number is the one in the source, not a general switchboard the model has seen on other letters.

The officer's review is a citation walk, not a style contest. Fluency issues get edited. Unsupported clauses get cut or left blank. Only then does the officer send.

Keep send authority outside the generator. The same person (or rota) who would have signed the English letter signs the translated one. Channel rules do not change because the text was drafted by a model. If SMS requires a character cap and an opt-out line that the locked source already includes, those constraints still apply. If they were never in the source, do not invent them at send time either.

Record what was sent, in which language, against which source version, and by whom. That log is what you produce when a resident says they were told something the English notice never said.

If the recipient later asks a follow-up that the notice does not answer, do not extend the generator into a live Q&A on the same thread. Hand that to the process you use for questions, including a citizen-facing policy assistant that is bounded to published policy rather than to a single outbound letter.

Outbound alerts that a resident may be due something they have not claimed are a different product. Those belong with proactive entitlement alerting, which still must not let a translator invent the entitlement itself.

A notice generator restates one locked English artefact. A public-facing data summarizer compresses published datasets for a general audience. Mixing the two is how a collection letter quietly picks up a citywide average wait time that was never in the resident's booking.

Keep eligibility, translation, alerting, and public explanation on separate processes, even if they share a vendor platform. The communications officer owns whether the words match the source. The service team owns whether the source was true. The model owns neither decision.

Most programmes already have a translation or messaging layer somewhere: Microsoft productivity and cloud stacks, Google Translate-class tools for raw language conversion, Salesforce Government Cloud for case and resident records, CivicPlus for public-site and resident-facing content. Treat those as a class of places a draft might be stored, routed, or displayed. None of them, by sitting in the architecture, make an uncited translation safe to send.

Whatever the platform, the control is the same. Lock the English source. Translate with cites. Leave blanks. A human sends.