One overall scoring question, or separate judgments combined with weights in code?

TypeSafe's Composite scoring documentation proposes splitting a complex judgment into independent dimensions, scoring them separately, and then normalizing and applying weights in code. This pattern lets developers inspect what contributes to the combined score. It describes a way to build a system, not an experimental finding that adding several scores necessarily improves accuracy. Points to debate The editors suggest considering a content recommendation example: relevance, actionability, and timeliness can each be defined separately, followed by explicit rules for combining them. If all of these are folded into one question—"Is this article worth recommending?"—the team may struggle to identify which criterion caused a ranking change. Splitting the judgment also has costs. Dimensions may overlap: "complete content" and "sufficient explanation," for instance, can reward the same feature twice. Some conditions should be disqualifying rather than offset by high scores elsewhere. The weights themselves express product values and should not be presented as objective facts discovered by the model. One approach is to check the combination rules against boundary cases that humans have accepted or rejected, then see whether small changes in weights produce large ranking reversals. If they do, discuss what the business can tolerate rather than simply picking weights that produce attractive results. Discussion questions 1. Which dimensions of your overall judgment could be annotated separately? 2. Which conditions must be hard gates that other scores cannot offset? 3. After changing weights, what fixed set of examples would you use to check that the ranking still makes sense? Primary source TypeSafe: Composite scoring https://docs.typesafe.ai/patterns/composite-scoring Editorial note: This official S1 Hub editorial account used AI to help compile public materials. This is not a submission by the original authors, and S1 Hub has not independently tested the work described. Additional evidence and different perspectives are welcome.