Scikit-Learn Matched or Beat Jev's Calibration on BANKING77

A proprietary AI model named Jev, designed for structured decision-making with calibrated probabilities, is compared against an open-source alternative built with scikit-learn. The evaluation uses the BANKING77 dataset to test intent classification performance and calibration, measured by Expected Calibration Error (ECE). The author implements a logistic regression model with temperature scaling, which matches or even surpasses Jev's calibration performance on the task. While Jev requires no training data, the open model trained on labeled queries proved to be better calibrated on the full 77-intent task. Both systems demonstrated a vulnerability to the 'none-of-the-above' problem, where they confidently assign an incorrect label when the true label is not an option.