DeciShift: Explain why decisions changed between ML system versions

Abstract

DeciShift v0.3.1 — Trust and evaluation hardening This release strengthens the trust layer around decision-change evidence while adding reproducible public-data evaluation. Highlights Public trained DecisionFlow examples using scikit-learn Wine and XGBoost Digits datasets. Machine-generated decision-change case study: improved model metrics while 4.17% of individual actions changed, with exact component attribution and a governed cohort contract returning BLOCK. Commit-tied 10K/100K exact-vs-sampled attribution benchmark snapshots with cache reuse, wall time, Python allocation and convergence diagnostics. Legacy DecisionPipeline caches bound to input-content fingerprints with defensive-copy and caller-input isolation protections. Hybrid replay invalidation when record content changes. Enforced branch-aware coverage floor raised to 80% with targeted trust-path regression tests. GitHub Actions refreshed and pinned to immutable commit SHAs; release publication continues through PyPI Trusted Publishing/OIDC. Expanded security and reproducibility documentation, including integrity-vs-authenticity limits. Release-source note The v0.3.1 tag is created from exact commit 082250abc1c0a82a41480949594168bb56f323ad, which passed the repository's full post-merge test matrix. The repository owner explicitly authorized this release without main branch protection; issue #7 tracks enabling branch-level enforcement for future releases. Scientific scope DeciShift attribution is software-counterfactual attribution. It does not establish real-world causality, safety, fairness, compliance or production fitness.