Open-source research · 2026
GovLLM
Policy-as-code
Legal requirements become rules a machine can check. An AI system declared compliant once, at launch, keeps changing; GovLLM checks every response against those rules, while it runs.
Problem
How do you justify a model choice six months after go-live? One-off benchmarks cannot answer, and the AI models used as judges have biases of their own.
Approach
Governance by design: each use case gets a profile written as code, up to 14 criteria mapped to the AI Act, GDPR and ANSSI. Small local models judge every output against it, and each request goes to the model that meets its rules.
Also
- OECD.AIlisted in the Catalogue of Tools and Metrics, June 2026
- arXivpreprint “Who judges the judges?”, May 2026
- 30 ★on GitHub, and discussed on Parlez-moi d’IA n°106