Use Python 3.10 on Linux for the recorded reference environment. Supported
dependency ranges are in requirements.txt; the historical snapshot is in
requirements-lock.txt. The setup helper creates .venv/ without changing
system Python:
PYTHON=python3.10 bash scripts/setup_environment.sh
source .venv/bin/activateAdd --locked to install the recorded dependency snapshot. The environment
also includes the Hugging Face download client.
Copy .env.example to .env for your model endpoints. Copy .env.mcp.example
to .env.mcp and supply only the credentials needed by the servers you enable.
Do not commit either local file. Model names in historical records identify the
experiment; availability of the original provider build is not guaranteed.
Live MCP execution additionally needs Node.js (22 or newer recommended) for
Node adapters, and uv for isolated Python servers:
python -m pip install uv
python mcp_servers/install.py --install-dependenciesSeven upstream projects have no explicit redistribution license and their
sources are omitted. Inspect THIRD_PARTY_NOTICES.md before explicitly opting
into their pinned-source fetch:
python mcp_servers/install.py --fetch-external --accept-unlicensed-upstream
python mcp_servers/install.py --install-dependenciesConstruct new records: follow the four commands in the root README. Profile
generation defaults to artifacts/profile_runs/latest/; the scenario and
trajectory commands pass that profile path explicitly. Outputs are local and
ignored by Git. Review each stage before proceeding.
Reproduce frozen results: obtain approved external data using
docs/data_access.md, then run:
python -m eval.reproduce_results --input-root .If data were extracted elsewhere, pass that directory as --input-root. It
must contain artifacts/profile_pool/, artifacts/scenario/,
artifacts/trajectories/, and artifacts/privacy_experiments/.
Generating new model outputs is not equivalent to reproducing the frozen results: provider builds, tool responses, and sampling can change.