LLM workflows and agents depend on binary choices, such as allowing or blocking an input, or discarding or retaining a retrieved document. In many settings, there is no option to defer to a human reviewer, so even uncertain cases must be assigned an action. A new arXiv paper (2610.19811) addresses this problem by studying System One models in the context of calibration and risk-budgeted decision-making.

Calibration ensures that predicted probabilities reflect true outcomes, while risk budgeting allocates allowable error across decisions. The authors appear to apply these principles to System One models so that automated binary choices remain reliable even without human oversight. The abstract is brief, so the exact method is not fully detailed, but the focus is clear: making forced binary decisions in LLM systems more trustworthy.