Research

Open research for a multilingual region.

We study how AI behaves in Arabic, its dialects and French, and whether its answers stay tied to evidence. Three technical notes set out our engineering perspective.

Our first question Browse the technical notes

Featured technical note

RN / 01

Sovereign architecture

Sovereignty is a system property, not a hosting label

Local model hosting is only one part of institutional control. Data flows, external dependencies and operating authority matter too.

Read the note

Open research · Question 01

Our first question

Does an AI document reader keep the same evidence when the same question is asked in French, Modern Standard Arabic or Moroccan Darija?

The question is what changes when only the language of the question changes. Moroccan Darija is the first dialect we study.

Three languages, one meaning?Mechanism illustration · illustrative values · no live AI

Françaisالعربيةالدارجة

Mechanism illustration with illustrative values, not results; no live AI. The same question, asked in French, Modern Standard Arabic and Moroccan Darija, becomes three points in one map of the document’s passages. All three retrieve § 2, with illustrative scores of 0.85, 0.82 and 0.76, the dialect a little further away. Each answer, in its own language, cites § 2. They are compared on answers that change, citations that do not support the answer, and missed or excessive refusals.

Illustrative values, not results

Each dot is a passage, placed by meaning in a 2D projection: closer means more similar.

What we look at

Answers that change
Does the answer itself change with the language of the question?
Citations that do not support the answer
Does the cited passage actually say what the answer claims?
Missed or excessive refusals
Does the system decline when the document contains the answer, or answer when it does not?

How we publish: When we publish, we publish the data, code, results and limits together.; This is a research question, not a published result.
What is published today: Three technical notes, below. They are engineering perspectives, not experimental studies.

Technical notes

Three questions about the whole system.

Read each note on its own or in sequence. Each is an engineering perspective, not an experimental study.

Sovereign architecture

Sovereignty is a system property, not a hosting label

Local model hosting is only one part of institutional control. Data flows, external dependencies and operating authority matter too.

Read the note

What remains outside the institution’s control?

Arabic and multilingual AI

Evaluating multilingual AI in institutional context

General benchmarks rarely capture the language mix, terminology, workflows, and consequences of a real institution.

Read the note

Does the system work in the languages of the task?

Private inference

Private inference in constrained environments

Restricted connectivity, limited hardware, and strict operational boundaries change how an AI system should be designed and maintained.

Read the note

Can the system be operated within its constraints?

From questions to evidence

A claim needs a way to be tested.

The same discipline runs through these notes.

State the conditions

Which task, language, data, infrastructure and access boundary does the claim apply to?

Look beyond the model

What do retrieval, permissions, external services and human review change about the result?

Keep the limits visible

What has been evaluated, what remains uncertain and what changes would require another assessment?

Questions about this research?

Explore the technical approaches behind it, or write to us about a related question.

Explore our technology Contact MX4 AI