Auditable Decision Models with Learned Abstention and Real-Time Steering

In three linesEvaluatorDPT is a bounded decision-control model predicting YES, NO, or TBD (learned deferral). Using a transformer encoder with structured auxiliary heads, it achieves Accuracy=0.8260 and Macro F1=0.8252 on 44,597 test samples. The interface enables inspectable routing and auditable decision control for production AI systems.

Summary generated by Claude — human-verified