DeciShift: Explain why decisions changed between ML system versions

DeciShift v0.2.0 adds a Trustworthy Decision-Change Evidence layer while preserving the v0.1 public API and CPU-first, local-first design.

Added

• Streaming Monte Carlo uncertainty for approximate component attribution, including standard errors and configurable confidence intervals.
• Attribution diagnostics with Shapley efficiency residual checks and warnings.
• Stable component provenance and strict reproducibility mode.
• Evidence schema 1.0, deterministic fingerprints, manifests, SHA-256 integrity verification, and decishift verify.
• User-defined Decision Contracts with deterministic decishift gate exit codes for CI.
• Stronger record identity, finite-output, row-order, and callable-invocation validation.
• Dynamic YAML threshold factories/objects while retaining numeric thresholds.
• Safer cohort discovery with Wilson intervals, coverage, global context, and excess flip rate.
• Optional binary outcome correctness-transition analysis.
• Decision Fragility diagnostics for threshold proximity and boundary crossing.
• Fully local self-contained HTML reports.
• Saved-run comparison with decishift compare-runs.
• Expanded benchmarks, regression coverage, Ruff, build, and Twine checks.

Compatibility

Existing v0.1 public imports remain supported, including DecisionPipeline, compare_pipelines, compare_predictions, exact_attribution, approximate_attribution, and pairwise_interactions.

Install

pip install --upgrade decishift==0.2.0
decishift demo

Scientific scope

DeciShift attributes observed software-output changes across a versioned executable decision pipeline. Component attribution is software-counterfactual attribution and does not establish external real-world causal effects. A Decision Contract PASS means only that the user-configured checks passed; it is not proof of safety, fairness, compliance, or correctness.