Output variance is the degree to which generations differ for the same task. When results are inconsistent, treat the variation as a diagnostic signal: check whether the prompt leaves room for multiple valid interpretations before blaming the model or API. section 1.6.3 section 8.1.1
Use these checks to locate and manage variability:
- Look for underspecification. Add an output schema, constraint, or worked example to narrow the prompt’s possible interpretations. section 1.6.3
- Consider sampling settings as a separate cause. Temperature 0 minimizes variability; moderate temperature permits useful variation, while very high temperature can increase noise. section 8.1.1
Support-ticket classification
If classification works when tickets use internal feature names but varies when users describe features informally, add a mapping from common descriptions to the internal taxonomy. This closes an interpretive gap in the prompt. section 1.6.3
