Abstention

A behavior in which a system avoids answering, deciding, or acting when evidence, estimated competence, or safety conditions are insufficient. The output may request more context, route the case to a person, or state that there is no basis for a conclusion.

Abstention requires a measurable policy. An overly conservative threshold reduces coverage; a permissive threshold preserves risky answers. Evaluation should track selective risk, coverage rate, and escalation cost on representative cases because a model's stated confidence is not calibrated by default.