Trust calibration for agentic tool use as preference learning: a GP-probit allow/ask/block policy gateway framed as Preferential Bayesian Optimization, with the paper and a reproducible simulation. - View it on GitHub