If confidence is 0.9, can you display “a 90% chance of being correct”?

TypeSafe's official documentation explains that confidence for Choice and Score is a statistic calculated from the returned probability distribution, not a separate round of independent verification. The two question types also use different calculations: Choice measures how much the highest probability stands out relative to a uniform distribution, while Score also considers distances across ordered levels. Sharing a field name therefore does not mean the value directly equals accuracy on your business data. Points for debate The editors suggest recording “how concentrated the model's distribution is” separately from “how accurate predictions of this kind actually are.” The former comes from one response; the latter requires a set of examples with reliable labels. If an interface presents a value as a percentage likelihood of being correct, it should first explain how that number was validated, rather than merely relabeling a field. Thresholds should not be copied mechanically between different questions, either. Changes in the number of options, level definitions, or input sources all warrant another check. One global cutoff is easier to manage, but may be too permissive for one type of input and too strict for another. We would prefer to see error cases and coverage presented by task, rather than just a high average confidence value. Let's discuss 1. Do you currently store the original probability distribution, confidence, or only the final label? 2. Does your product present the concentration of model outputs as validated accuracy? 3. Should different questions share a threshold, or have separate acceptance criteria? Primary source TypeSafe: Confidence https://docs.typesafe.ai/confidence Editorial note: The official S1 Hub editorial account used AI assistance to compile public sources. This was not submitted by the original author, and this site has not conducted independent hands-on testing. Additional evidence and differing views are welcome.