A consultation rarely arrives in one language. This is what it looks like when Prism reads the same submissions across fourteen of them — recognising the same claim wherever it appears, and citing each one back to the exact words the citizen wrote, in the language they wrote them.
To test cross-language reading honestly you need documents whose meaning is known to be identical across languages — otherwise "did the system recognise these as the same claim?" has no definite answer to check against. So we took three real submissions to the Irish Citizens' Assembly on Gender Equality and had them carefully translated into thirteen further European languages.
The translation is the only artificial part, and it exists purely to construct a known answer. The submissions are real; the point being tested — whether Prism reads each language and lands on the same claim — is real. And the arrangement is deliberately conservative: any imperfection in a translation makes the system's job harder, never easier, so a clean result cannot be an artefact of a helpfully tidy translation.
Prism reads each submission, distils it into short claim phrases, and groups those phrases by meaning. The question this example poses is simple: does each claim gather its fourteen language versions into one group, or does any group instead fracture along language lines — the Finnish in one place, the Greek in another?
No group split by language. Every distinct claim gathered its language versions together, even when the grouping was made fine enough to separate closely-related claims within a single submission. And the measure that matters most: reading a claim in a different language moved it no more than re-reading the very same text would. The spread introduced by fourteen languages was no larger than the system's own run-to-run variation on identical text. On this evidence, the analysis is robust to the language a citizen happens to write in.
It would be natural to assume the system first translates everything into English and works from that. It does not, and that distinction is the whole point. Prism reads the submission in its original language, reasons about it in English, and — critically — quotes the citizen verbatim, in the language they wrote. Nothing the citizen said is ever paraphrased through a translation before an analyst sees it.
What makes that safe rather than reckless is grounding. Every claim the system reports is tied back to the exact stretch of characters it was drawn from, in the source text. If the reading of a French or Spanish submission is sound, the proof is the French or Spanish sentence sitting under the claim, unaltered. Across the 126 readings in this test — three submissions in fourteen languages, each read three times — every quoted span was located in its own source text, with none left unresolved. The examples below are exactly that: a single claim — that councillors should be paid enough for ordinary people to afford to stand for election — read from four widely-spoken languages, each shown with the original words it was grounded to. English comes first, as the real submission text.
“We call for an increase in the salary of County and City Councillors to the average industrial FTE salary … those including women who chose to enter public life and stand for election will have access to a living wage.”
„Wir fordern eine Anhebung der Gehälter der County- und City-Councillor auf das average industrial FTE salary … Zugang zu einem existenzsichernden Lohn haben.“
« une augmentation de la rémunération des Councillors de comté et de ville pour l'aligner sur l'average industrial FTE salary … aient accès à un salaire décent »
«Pedimos un aumento del salario de los Concejales de condado y de ciudad hasta el average industrial FTE salary … tengan acceso a un salario digno.»
Four independent readings, four slightly different phrasings — “salaries” and “pay”, “living wage” and “decent salary” — landing on one claim, each pinned to the words actually written.
Widely-spoken or not, the same holds — and “your own language” means everyone's, not just the largest. Here is one more claim, paid maternity leave for elected representatives, read from Hungarian, from Finnish, and from Greek: two languages with grammar heavy enough to be chosen as deliberate hard cases, and one written in an entirely different alphabet.
„a megválasztott városi és megyei Councillorok, TD-k és Senatorok azonnal hozzáférhessenek fizetett szülési szabadsághoz”
”Vaadimme lisäksi välitöntä oikeutta palkalliseen äitiysvapaaseen vaaleilla valituille kaupungin- ja kunnanvaltuutetuille (Councillor), TD:eille ja Senatoreille”
«Ζητούμε επιπλέον την άμεση πρόσβαση σε αμειβόμενη άδεια μητρότητας για τους εκλεγμένους Councillors Πόλεων και Κομητειών, τους TD και τους Senators»
Here the phrasing did more than land on the same claim — all three came out identical, word for word, though each was read independently from a different language and script. That exact agreement is not something the system is told to produce; it is what happened. The wording a large language model produces varies a little from reading to reading, as the four languages above show. What does not vary is where each phrase is anchored, and which claim it belongs to.
This example is built to rule out a specific failure: that reading many languages quietly fragments the analysis along language lines. It doesn't — the languages stay commensurable, and every claim keeps its citation in the original. That is the property a multilingual consultation most needs and most easily loses.
The Prism overview shows the whole approach. If you work with consultations that arrive in more than one language and need them read without flattening them into translation, we'd be glad to walk you through it.
info@szlamkaconsulting.comSource: PRISM synthetic multilingual corpus · clustering run 20260714-223010 (k=35). The three source submissions are real Citizens' Assembly filings; the thirteen non-English renderings are machine translations, generated only to construct a known-answer test. Citations shown are the verbatim grounded spans. Figures reproducible from the same run.