Release and benchmark evidence#
Level: Expert
You will: produce bounded, reproducible evidence and interpret release and
benchmark claims without overstating hardware capacity.
Before you start: Verified capabilities.
Services: None; GitHub and PyPI publishing are release-time services.
A release artifact is trustworthy only when the code, documentation, package,
and evidence agree. Pull-request CI runs the test suite before merge. The
release workflow is activated only by a vX.Y.Z tag; it validates the package
metadata and Avro contracts, verifies the built wheel in a clean environment,
builds strict documentation, and attaches the wheel and source distribution to
the GitHub release before trusted PyPI publication.
Laptop evidence protocol#
poetry run fraudtwin scale-benchmark configs/scale-dev.yaml \
--output-dir runs/evidence-dev \
--checkpoint-dir runs/evidence-dev-checkpoint \
--evidence-dir benchmarks/evidence
Commit the resulting evidence only when it contains the resolved configuration
hash, seed, Git revision, package versions, host summary, output fingerprints,
resume result, and explicit laptop-dev-only claim scope. Never describe this
result as a 100M/1B capacity measurement.
Benchmark-pack metrics and scale evidence answer different questions: packs compare deterministic fraud/ML conditions; scale evidence describes execution and storage behavior. Keep their manifests and claims separate.
Next#
Use Release readiness to review what is supported, or follow Development to reproduce the checks locally.