Run these gates before publishing a public release.
pytest -q
PYTHONPATH=. python tests/run_smoke_tests.py
PYTHONPATH=. python examples/acme_demo.pyPYTHONPATH=. python -m benchmarks.decision_risk.run \
--n 1 \
--budgets 40 \
--distractors 5 \
--output-dir outputs/release_verify_decision
PYTHONPATH=. python -m benchmarks.decision_risk_human_rewrite.run \
--n 1 \
--budgets 40 \
--distractors 5 \
--metadata-noise 0.18 \
--output-dir outputs/release_verify_rewrite
PYTHONPATH=. python -m benchmarks.decision_risk_stress.run \
--n 1 \
--budgets 40 \
--distractors 50 \
--metadata-noise 0.30 \
--output-dir outputs/release_verify_stress- README says the benchmark is synthetic.
- README says generated-answer scoring is deterministic and label-aided.
- README does not claim production proof or real LLM preference wins.
- CAC core and non-oracle baselines do not read gold scorer labels.
- Large archive zips are published as release assets, not committed to git.