In July 2026, shipped a critical bug fix in mlflow/mlflow addressing an issue where binary classification evaluation curves ignored the user-specified positive label and relied on a hardcoded column index. The fix introduces dynamic column selection based on the configured positive label index and adds a regression test to ensure AUC metrics respect the configured positive class. This change reduces the risk of misinterpreting model performance and improves reliability of ROC/PR/calibration metrics for binary classifiers.

June 2026 monthly summary: Delivered Unified Reset and Refit Capability for Conformal Models in MAPIE, enabling safe refit and data replacement across CrossConformalRegressor, JackknifeAfterBootstrapRegressor, and CrossConformalClassifier; includes state management warnings when refits occur without a reset and accompanying tests. These changes improve reliability of conformity scores, reduce risk of stale predictions, and accelerate model iteration.

Concise month-end summary for MAPIE (2026-05): Focused on stabilizing the Prediction Interval Optimization workflow to deliver reliable interval estimates in regression conformity scores. Implemented a critical bug fix to ensure optimization actually minimizes interval width, and expanded test coverage to guard against regressions. Result is a more trustworthy, production-ready optimization routine with better alignment to downstream decision thresholds.