Skip to content

Reproducible reference study

frsutils includes a real-dataset research artifact under studies/fuzzy_rough_reference_study/. It provides inspectable evidence that the package-root API can support a downstream scientific workflow.

The study applies ITFRS, VQRS, and OWAFRS to three binary tasks derived from public scikit-learn datasets. It records per-sample approximation values, verifies dense/blockwise numerical equivalence, measures repeated runtimes, runs the repository benchmark profile, captures the software and hardware environment, and writes checksums for generated outputs.

Run the study

python studies/fuzzy_rough_reference_study/run_study.py

The canonical configuration is:

studies/fuzzy_rough_reference_study/study_config.json

Detailed methods and the committed result snapshot are available in:

studies/fuzzy_rough_reference_study/README.md
studies/fuzzy_rough_reference_study/results/

Scope of the evidence

The artifact supports the following claims:

  • one stable package-root API executes all three documented model families;
  • exact blockwise NumPy execution reproduces dense NumPy outputs within the configured tolerance;
  • configurations, per-sample outputs, runtimes, package versions, platform metadata, and checksums are machine-readable;
  • no external dataset download is required.

It does not claim that one fuzzy-rough model is universally superior, that runtime values generalize to other hardware, or that Python-level memory measurements capture every native allocator.

The committed snapshot includes environment.json, which records the package version, Git commit, worktree state, Python environment, and platform metadata. A clean provenance record reports git_worktree_dirty: false.