# Audit a Bilingual Persona: Check Facts Across Locales

Use a bilingual audit grid to find added facts or unknowns changed during persona localization.

- Canonical URL: https://www.harmate.com/en/blog/audit-a-bilingual-persona-check-facts-across-locales
- Author: Harmate Team
- Published: 2026-11-15
- Updated: 2026-11-15T01:55:14.872901+00:00
- Language: en

## Content

# Audit a Bilingual Persona: Check Facts Across Locales

Translating a test persona can silently change what an assistant is meant to know. A delivery date added in one language, an unknown turned into a confirmed fact, or a different rule can change the scenario. A bilingual audit checks that facts and their states match. It does not assess general model quality or the effect of language on model responses.

The example below is a fictional editorial check created for this article. It does not describe a profile from Harmate’s persona bank.

## Compare the fields one by one

Start with a canonical card that records the expected data and states. Create each localized version with the same structure. For every field, compare its value, role, and state: confirmed, unknown, absent, or prohibited from inference.

| Canonical field | French | English | Audit result |
|---|---|---|---|
| Confirmed status: “In transit” | En transit | In transit | Equivalent |
| Latest event: November 14, 2026 | 14 novembre 2026 | November 14, 2026 | Equivalent; this is the latest event |
| Delivery date: unknown | Non confirmée | November 18, 2026 | Mismatch: English adds a confirmed date |
| Address: unknown | Non renseignée | Not provided | Equivalent |
| No return window supplied | Aucun délai communiqué | Return within 30 days | Mismatch: English adds an absent policy |

This example contains two mismatches: an invented date and an added return policy. That count describes only the two errors deliberately inserted into this fictional pair; it is not a translation error rate.

## A reusable review procedure

For each pair, follow these steps:

1. Freeze the canonical card and its version before translation.
2. Align every field, including unknowns and no-inference rules.
3. Compare facts and dates without confusing an event date with an expected date.
4. Check that both locales have the same facts, constraints, and missing information.
5. Record each mismatch, correct it from the canonical source, then review the full pair again.

In this example, correction means removing November 18 and the 30-day rule, then stating in both languages that the delivery date is unknown and no return window was supplied. Repeat the comparison after editing; fixing one sentence without reviewing the rest could leave another mismatch.

Wording can vary naturally: date formats, syntax, and local expressions do not need to match word for word. The facts and their states do need to match. Keep the audit table with the card, localized versions, review date, and reviewer. This is a content consistency check; it does not evaluate language proficiency or model outputs.

Harmate’s page about [personas](https://www.harmate.com/en/pricing/api#personae-example-title) contains a public example. The table in this article is a separate illustration, created only to demonstrate the audit method.

After correcting a translation, use the [test-set diff](https://www.harmate.com/en/blog/review-test-set-changes-with-a-readable-diff) to see which cases changed. The [versioned regression case](https://www.harmate.com/en/blog/build-an-ai-agent-regression-test-with-versioned-persona-data) shows what to hold fixed when replaying them.