Meta open-sources MCGrad for web-scale multicalibration

MCGrad shifts multicalibration from a theoretical "fairness" constraint to a high-performance production requirement that actually improves model accuracy.

• –Automated Subgroup Discovery: It finds miscalibrated regions mathematically using residuals, eliminating the need to manually define every possible feature intersection.
• –Performance Boost: Meta reports that fixing subgroup calibration improved Log Loss and PRAUC on 88% of production models, proving that reliability and accuracy are not a zero-sum game.
• –Web-Scale Scalability: Designed to handle over a million real-time predictions per second, it is far more practical for large-scale industrial use than previous research-stage algorithms.
• –Low-Overhead Integration: The scikit-learn compatible API and use of optimized libraries like LightGBM make it easy for teams to bolt onto existing pipelines with minimal friction.