In the repository

Run it with pnpm harness from the repo root; paths above are under harness/.
The tester's llama.cpp: docker compose up -d llama (see tester/README.md; port 8081, LLAMA_PORT=... to move it). Only run models from the publishers in tester/model-allowlist.txt (ADR-0004).
For a server on your machine use --server http://localhost:8081.
Why the harness looks like this: ADR-0019 and the ADRs it links to.