Confidence and visible evidence

Calibrate the operator's confidence unit and preserve useful visible reasons.

Routing decides which work reaches a semantic operator. Confidence decides how the operator's returned evidence is admitted. These are separate contracts. A confidence score describes the strength of evidence for one complete decision. The task's confidence_threshold controls acceptance; changing it does not change the meaning of the score.

A useful rubric defines what one score certifies, which uncertainty weakens it, and examples of well-supported and uncertain decisions. A model's score is not a measured accuracy guarantee. With visible reasons enabled, length guidance controls generation only; Relifold never truncates an already generated reason.

See Cost, semantic review, and confidence for the complete operator-specific table and worked confidence example.