Reliable Defect Probabilities Across Software Releases: A Leakage-Controlled Calibration Study - Reproducibility Package

Résumé fourni par la source

Reproducibility package for Reliable Defect Probabilities Across Software Releases: A Leakage-Controlled Calibration Study. The original experiments cover 9 Java projects, 32 releases, 65 metrics, 14 chronological release transitions, four classifiers, and 56 project-transition-classifier tasks. The archive contains original Python source code, fixed configurations, tests, protocols, portable provenance metadata, and aggregate results. Version 1.1.0 adds the September 2026 exploratory temperature and beta calibration baselines, calibration-harm analyses, and a historical-only method-selection analysis. The original six-method analysis is retained separately from these exploratory extensions. The package includes 1,008 original metric rows, 1,680 extension metric rows, and 672 optimizer-fit records. Historical selection uses previously observed outcomes, with current target outcomes reserved for evaluation. Reproducibility limitation: the recorded extension run failed its strict all-metric reproduction check. The failure and a separate diagnostic are retained. The maximum probability difference was 1.22e-13, and continuous fields matched the fixed 1e-10 tolerance, but exact ranking ties and five threshold memberships changed. Exact all-metric replication is not claimed. The extracted package passed 31 tests; two benchmark-dependent tests were skipped because third-party data are not included. This archive excludes third-party benchmark rows, sample-level or file-level predictions, private out-of-fold arrays, credentials, caches, environments, manuscripts, cover letters, and journal templates. Obtain the Large Defect Prediction Benchmark v1.0 separately from https://doi.org/10.5281/zenodo.6342328 at Git commit 0d40a69f047d4f508f92e196b514a47bfa653da1 and review its reuse terms. Original software is distributed under the MIT License; this license does not relicense third-party datasets, dependencies, or other third-party materials. Previous version: https://doi.org/10.5281/zenodo.21508396. Concept DOI for all versions: https://doi.org/10.5281/zenodo.21508395. File: supporting_code.zip (1,002,950 bytes). SHA-256: c436cbcc6497364fc30405b8798536cb54f2958c63ae6626fa469755fb80ea12.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.