A confidence figure attached to an automated score tells you where the evidence was clear and where it was thin. A disclosure either was or was not read out, and a model should be near-certain either way. Whether an agent showed genuine empathy is a judgement call, and a low confidence figure on that parameter is the honest answer.
Used properly, confidence is a routing signal: it tells you which calls deserve a human ear. Used carelessly, it is another number to average.
It is a self-report. A model can be confidently wrong, and no confidence figure changes that.
Every parameter comes back with a 0–100 confidence figure, and the call carries the average. Where the model reports none, we show a dash rather than a default: a made-up confidence would defeat the entire purpose of printing one.
Evidence & confidence