Auditable Decision Models with Learned Abstention and Real-Time Steering
EvaluatorDPT, introduced in arXiv paper 2605.27768, is a bounded decision-control model that explicitly handles uncertainty by predicting YES, NO, or TBD (to be determined). Unlike forced classifiers that collapse uncertain cases into action labels, TBD is learned as a deferral outcome, not added post-hoc. The model uses a transformer encoder with a primary bounded-decision head and structured auxiliary channels for values and emotions/sentiments. It is domain-agnostic: a deployment domain supplies evidence and policy thresholds, while the model emits a bounded distribution controllable at inference time through recorded operating thresholds and validated auxiliary semantic signals. The paper reports decision performance on held-out validation sets, though specific metrics are not detailed in the excerpt. This approach enables auditable, policy-governed decision-making, allowing developers to route uncertain cases to human review or alternative processes rather than forcing a prediction.
Enables auditable deferral of uncertain AI decisions to human review.