Reproducibility

dfsc separates installation verification from paper-experiment reproduction.

Installed-artifact verification

Create a new environment, install the released wheel from PyPI or the GitHub Release, and run:

python reproduction/verify_install.py

The script records the package version, operating system, Python and PyTorch versions, accelerator information, and CPU/GPU forward-backward smoke tests. It exits with a nonzero status if the installed version is not the expected release or if any value or gradient is non-finite.

For the 0.1.0 release, install from PyPI:

python -m pip install "dfsc==0.1.0"

The archived 0.1.0rc1 TestPyPI artifact and its verification hashes remain recorded in reproduction/release_manifest.json.

Source-tree verification

From a clean checkout of the matching tag:

python -m pip install ".[test,docs]"
python -m unittest discover -s tests -p "test_*.py"
python -m mkdocs build --strict
python -B tools/pre_release_audit.py
python -B experiments/exp35_software_artifact_audit.py

Core maturity experiments can then be run with:

python tools/reproduce_core.py

Some real-data experiments require the source datasets described in their manifests. The AnomDiffDB raw files are not redistributed until their redistribution terms are explicit.

Independent report

An independent verifier should start from the public release rather than a developer environment and complete reproduction/EXTERNAL_REPRODUCTION_REPORT.md. A GitHub-hosted CI run demonstrates clean infrastructure execution; it is not labeled an independent human reproduction.