Open Benchmarks for Civic LLMs

This publication proposes a benchmark suite for evaluating language models used in public services.

We emphasize:

• failure modes that matter for citizens,
• documentation that procurement teams can read,
• and audit trails that make evaluation reproducible.