Measured question kinds
Each row is one kind of question, asked up to 400 times on recordings from sources Sro never trained on. We publish where it is weak as well as where it is strong.
Calibration error is the gap between the probability Sro states and the share of those answers that turn out right. A low number means you can set a threshold on Sro's answer and get the error rate you expect.
The API attaches these measurements to every matching question you ask. A question outside this list still gets an answer, marked as unmeasured.