v0.1.16

Highlights

• Adds paired binary-classifier and reference-group comparisons, probability calibration and decision metrics, conditional thresholds, numeric long-form exports, and explicit inference configuration.
• Adds Wilson and exact interval policies alongside diagnosed percentile resampling, with preserved sample point estimates and interval provenance.
• Adds class-split score distributions, dedicated interval plotting, hierarchy fixes, stricter validation, optional plotting/styling extras, and expanded statistical regression coverage.

Migration notes

• Zero-denominator rates, undefined ratios, and unidentified calibration fits now return NaN instead of a misleading zero.
• Evaluation tables are neutral by default. Opt in to relative metric ranking when desired; enrichment odds ratios are never performance-colored.
• split_classes=True now requires a registered binary outcome and overlays class histograms. The default remains False.
• Missing subgroup values, categorical ordering, invalid scores/truth values, duplicate metrics, and bootstrap diagnostics are handled explicitly. Review code that relied on permissive or implicit behavior.
• Core evaluation remains available without rendering dependencies. Install model-auditor[plotting], model-auditor[styling], or both as needed.

Full Changelog: v0.1.15...v0.1.16