When Choice gets it wrong, could the candidate set be the problem?

The official Choice documentation explains that it returns a selection from predefined options, along with a probability distribution. The option names and descriptions inform the model's judgment. The documentation also cautions that “other” or “none of the above” may not work well; a separate Noul question can assess whether the candidates apply before you decide whether to use the Choice result. Points for debate The editors suggest maintaining candidate sets as product rules. For example, “refunds,” “shipping,” and “billing” might all apply to one email. If the business needs a primary responsible team, the priority rule needs to be explicit. If it needs multiple labels, a single Choice question may already be the wrong formulation. Adding labels does not necessarily resolve overlap and may simply hide the problem in a longer list. One starting point is a small set of disputed real examples: have people explain why each belongs to a category, then put those boundaries into the descriptions. Keep a separate fallback path for inputs outside all known categories. We suggest tracking misclassifications, abstentions, and human workload together. Sending every difficult case to a person can yield impressive classification accuracy without actually reducing work. Let's discuss 1. Do your categories describe the content, or decide which team should handle it? 2. What explicit rule should determine the choice when two candidates both apply? 3. For an input that matches none of the candidates, would you prefer an “other” option or a separate applicability check first? Primary source TypeSafe: Choice https://docs.typesafe.ai/primitives/choice 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.