Changelog¶
This project follows semantic versioning where practical during alpha development. Alpha releases can still contain intentionally breaking API changes; pin exact versions.
0.2.0 — unreleased¶
Added¶
- Directed Bayes-factor helpers, pairwise tables, posterior model probabilities, independent-evidence aggregation, and Savage–Dickey/history effect utilities.
- Typed count/trajectory validation scenarios, recovery and coverage summaries, superiority/ROPE probabilities, and complete validation runners.
- Cumulative-prefix count and trajectory Bayes-factor scans with schema validation and replicate summaries.
- Donor-aware simulation, posterior extraction, variance decomposition, leave-one-donor-out mixture sensitivity, and independent-condition contrasts.
- SMC/posterior/trajectory diagnostics.
- Optional UI-neutral Matplotlib plotting.
- Atomic inference persistence and checksummed scan bundles/archives.
- MkDocs Material documentation and artifact-based GitHub Pages deployment.
Changed¶
README.mdis the canonical package and PyPI introduction.- Package-scale validation ceilings replace inherited web-demo limits. These are broad safety bounds, not practical workload recommendations; frontends should impose tighter limits.
- Package version advanced to 0.2.0.
Stability notes¶
- Version 0.2 remains alpha.
- Bayes-factor fields encode direction explicitly as
numerator_vs_denominator. - Scan sample sizes are nested cumulative prefixes within each scenario/replicate.
- Leave-one-donor-out moments recompute mixtures without refitting.
0.1.0¶
- Initial standalone package with event-count, donor-aware, condition-wise, and ordered-trajectory simulation/inference, result tables, and archives.