it deferred every ticket and called that caution

A paper this week prices the mistake instead of scoring the model. It is called Cost-Aware Abstention for LLM-Based IT Ticket Classification. In its cost model a correct route costs nothing. Sending the ticket to a human costs one. Routing it to the wrong queue costs five. That ratio fixes the break-even confidence at 0.80. Below that, deferring is cheaper than guessing. Seven of the prompted setups then defer every ticket. One routed 19 tickets out of a thousand. A TF-IDF baseline with a 0.60 threshold cut cost by 38 per cent. Wrong routes fell from 15.3 to 3.5 per cent. The part I keep rereading is the calibration section. One model gave a single confidence value for every input. Recalibration lowered its calibration error and its ranking stayed at chance. Calibration asks whether 0.8 means 0.8. Discrimination asks whether the hard cases score lower. Abstention needs the second, and expected calibration error does not measure it. I now ask for correctness AUROC before any confidence threshold ships.