Skip to content
Stentor

How STENTOR checks a translation

Machine translation exists. Agencies do not use it for alerts because a wrong alert is worse than no alert, and nobody can proofread ten languages under pressure. STENTOR does not ask anyone to trust a translation. It reads every translation back and only sends what still means the same thing.

The path of one alert

Worked example: the demo wildfire alert, Vietnamese.

  1. Step 1

    Read

    The English alert becomes an Alert Intent: hazard, required action, places, times, quantities, phone numbers and every instruction with a flag for “do not”.

    hazard wildfire · action evacuate now via P2 · NOT return until T1 · info U1

  2. Step 2

    Lock

    Names, streets, times, quantities and phone numbers are replaced by tokens before any translation, so no model can change them. Times are re-formatted per language on the way back, never re-valued.

    Wildfire near P1. Evacuate now via P2. Do NOT return until T1. Info: U1.

  3. Step 3

    Translate

    Approved phrase templates are used whenever they express the whole alert. Only what they cannot express goes to a model (an Adaption fine tuned model, then Claude), and that output is checked the same way.

    Cảnh báo cháy rừng tại P1. Sơ tán ngay theo đường P2. KHÔNG quay lại trước T1. Thông tin: U1.

  4. Step 4

    Read back

    Each translation is read back into an Alert Intent without seeing the English. With a Claude key this is a different model from the translator, so they do not share blind spots.

    hazard wildfire · action evacuate now via P2 · NOT return until T1 · info U1

  5. Step 5

    Compare

    The two intents are compared slot by slot. Any critical difference blocks the language; the operator sees exactly which slot failed and why.

    Hazard ✓ · Action ✓ · Places ✓ · Times ✓ · Phone/link ✓ · Negations ✓ · Steps ✓

  6. Step 6

    Send

    Verified languages go out as CAP 1.2 (one block per language), SMS, audio and a public page. A blocked language gets the pre-approved generic warning for that hazard instead of a wrong one.

    Cảnh báo cháy rừng tại Altadena. Sơ tán ngay theo đường Lake Ave. KHÔNG quay lại trước 18:00. Thông tin: 211.

The checks

All are critical: one failure blocks the language.

CheckWhat it catchesExample it blocks
Every literal exactly onceA dropped, duplicated or invented name, time or number.Model writes “until 7:00 PM” instead of keeping ⟦T1⟧ (6:00 PM).
Required action presentThe approved phrase for the action is missing.“Evacuate now” rendered as “consider leaving”.
No softer actionA weaker action replaces the required one.“Evacuate now” becomes “prepare to evacuate”.
Negation countFewer “do not” markers than prohibitions.“Do NOT use elevators” becomes “Use elevators”.
Slot by slot read backAny change of hazard, action, place, time, route vs destination, quantity, name or instruction.“Evacuate via Lake Ave” read back as “evacuate to Lake Ave”.
Literal restored verbatimA literal lost while putting the values back.Phone number 211 missing from the final text.

What each status means

✓ Verified

Every critical slot matches. Sent.

! Review

The meaning matches, but something needs a person: the templates leave out a word, or STENTOR had to assume the action. Not sent until someone decides.

✕ Blocked

A critical slot differs. Nothing in this language is sent.

↺ Safe fallback

The translation was blocked; the pre-approved generic warning for the hazard is sent instead, after passing the same checks.

Languages

Every language has an action lexicon, negation markers and about 30 approved instruction templates. All were drafted with an LLM and are marked unreviewed until a native speaker checks them.

Language“Evacuate now.”“Do NOT use elevators.”Negation markersReview
Start the API to see the live lexicons.

What STENTOR does not do

  • It does not replace language access staff. It makes their approved wording reusable and shows them exactly what needs them.
  • It does not judge style or politeness, only meaning in the slots it models. Words outside those slots send the alert to Review.
  • It cannot lock a place name written in lower case (“near altadena”). It flags it instead.
  • Without an API key the translator and the checker share one lexicon. That is safe for the templates it ships, but it is not an independent second opinion.
  • It is not certified for production alerting and does not connect to IPAWS or carriers. CAP files are marked as an exercise.

Adding or reviewing a language

  1. Copy services/api/stentor/lexicon/en.json to a new language code.
  2. Translate the hazard names, action sentences and instruction templates. Keep every {P}, {T} slot.
  3. List the negation markers and, for each action and instruction, a short phrase that always appears in its template.
  4. Run uv run pytest tests/test_lexicon.py: every template must read back to the same meaning.
  5. When a native speaker has checked the action phrases and negation markers, set reviewed_by to their name.