Worked Examples¶
The repository includes three small scripts that demonstrate the public Python API without blurring collection, evaluation, and presentation.
Reproduce Frozen Results¶
python examples/reproduce_frozen.py
python examples/reproduce_frozen.py --format markdown
This reads the archived canonical evaluation records and prints the same 21-model, 154-prompt summary used by the package. It does not execute generated code or contact a provider.
Build an Interactive Dashboard¶
python -m pip install ".[visualization]"
python examples/build_dashboard.py --output pqid-bench-dashboard.html
Use --plotlyjs inline when the HTML file must remain fully offline.
Plan a Live Model Run¶
python examples/plan_live_run.py \
--provider groq \
--model MODEL_ID \
--limit 3
The planner validates the route, evidence manifest, prompt selection, and
prospective output contract. It reports contacts_provider: false; no
credential or prompt-export acknowledgement is needed until an actual
run-model invocation.
Trust Boundary¶
| operation | contacts a provider | executes generated code |
|---|---|---|
| reproduce example | no | no |
| dashboard example | no | no |
| live-run planner | no | no |
pqid-bench run-model |
yes | no |
pqid-bench replay |
no | yes, isolated Docker worker |
See the complete examples directory, Live Model Testing, and Security and Governance.