How I built and trained a decision model that gives one calibrated probability for each option, in one forward pass.