S.A.F.E agentic: a trustworthy credit scoring system with a conversational platform

Abstract

Credit scoring is an area where Machine Learning models should be evaluated not only for predictive performance, but also for fairness, explainability, and robustness as part of a broader trustworthy-AI governance process. This paper presents a SAFE agentic credit-scoring system that coordinates multiple stages of Machine Learning evaluation and governance while retaining human oversight. The proposed framework includes data preparation, model training, evaluation, governance scoring, sensitivity analysis, and artifact-grounded reporting. Six candidate models are compared using 5-fold stratified cross-validation and extended hyperparameter optimization. The selected model is subsequently assessed using a SAFE governance score that combines rank-based accuracy (AURGA), reference-aligned Rank Graduation Robustness (RGR), rank-based explainability (AURGE), and a group-level Fairness Aggregate. Uncertainty is quantified across cross-validation folds, and the framework additionally evaluates calibration, distributional separation, and sensitivity to alternative governance weights. On the German Credit dataset, Random Forest achieved the highest observed mean model-selection ROC-AUC (0.8025) and was therefore selected for the subsequent SAFE evaluation. Its fold-specific SAFE evaluation yielded a mean ROC-AUC of 0.7977 and a mean SAFE Score of 0.8247 (95% CI [0.7850, 0.8644]). Fairness was the lowest and most variable SAFE dimension. Independent replication on the Taiwan credit dataset, using models fitted separately within 5-fold stratified cross-validation, yielded a mean ROC-AUC of 0.7808 and a mean SAFE Score of 0.8646 (95% CI [0.8564, 0.8728]). The weighting sensitivity analyses were additionally replicated on both datasets. The central contribution of this work is an end-to-end agentic governance prototype for credit scoring. It combines traditional Machine Learning methodology, structured SAFE assessment, rank-based metrics, uncertainty and sensitivity analysis, and artifact-grounded reporting to provide transparent and auditable evidence for human governance review.