Skip to content

Technology

Four approaches to AI that institutions can check and control.

Approaches

How we work on multilingual, sovereign AI.

Four practical questions: does the system work in the task’s languages, can its answers be traced, where does it run, where does its data come from?

Schematic

Multilingual evaluation

General benchmarks rarely reflect the languages of a real workflow. We treat Modern Standard Arabic, its dialects, French and English, and switching between them, as separate test conditions, together with the terminology of the domain.

Proposed method

Test questions per language · Code-switching cases · Domain terminology · Review by people who speak the language

Evidence-grounded retrieval

An answer is only as useful as the passage behind it. Answers should be tied to the passages that support them, citations checked against what those passages actually say, and a missing answer stated as such: “not in the document”.

Proposed method

Passage-level citations · Citation checks · Explicit “not in the document”

Private inference

Open-weight models can run inside an environment the institution controls, including constrained or isolated ones. Model choice follows the task, the languages, the licence and the hardware actually available.

Proposed method

Open-weight models · On premises or private cloud · Isolated operation

Governed language data

Evaluation and adaptation depend on language data whose origin and permitted use are known. Provenance, rights, versions and access rules should be documented alongside the data itself.

Proposed method

Documented provenance · Usage rights · Versioning

See where these approaches apply

Constraints

Control is a property of the whole system.

An open-weight model alone does not make a system sovereign.

Read our technical notes

Choose the model. Test its suitability.

Evaluate open-weight or approved private models against the task, language, licence and available hardware. Workflows in Arabic, its dialects, French and English need their own evaluation.

Define what can cross the boundary.

Specify where data is stored and processed, which services can communicate, and whether external endpoints are permitted. An isolated environment requires a different design from a connected one.

Make the system operable.

Define access administration, logging, updates and recovery procedures with the team responsible for running the system. Document the configuration and its dependencies.

Next

Questions about these approaches?

Read the research behind them, or tell us about a related project.