LLM based decision making
This page was generated with assistance from an LLM. The other pages on this site are handwritten.
A decision model does not write a reply. You give it a state and a few questions whose answers you already named. It returns a probability for every option, and ordinary code picks the action.
The three question types below are the ones used by Jev-style models: a choice among options, a yes probability (noul), and a score on an ordered scale. This page runs that pattern in your browser with a small zero-shot classifier. These probabilities are not from Jev, and they are not calibrated, so the cutoffs only show where code would branch.
Questions
- Choice. Which team should handle this? Returns, shipping, or billing.
- Noul. Does this need urgent human attention? A probability that the answer is yes.
- Score. How frustrated is the customer? Calm, frustrated, or very angry.
The classifier downloads when you press Decide.
One cutoff for extra care
A clinic offers a care-management program to patients whose risk score is at or above a cutoff. Patients below the line get routine care. These sixteen patients have the same kind of illness. The score predicts what the clinic will spend on them. Less is spent on Black patients who have the same need, so their scores sit lower.
The cutoff never reads the group. Because the scores sit in different places, that one line offers the program to fewer Black patients. These scores are written into the page so you can move the line. The same mechanism, a cost score and one enrollment cutoff, is what Obermeyer et al. reported in Science in 2019: at a given score, Black patients were sicker, and the cutoff identified fewer of them for extra care.