Ideas behind the experiment

Three nearby ideas

Authorization means deciding whether a subject may perform an operation on a resource under a policy. For example, a staff member may read a particular workbook only when permission covers that workbook. Attribute-based authorization considers properties of the person, resource, operation and environment.

Selective classification lets a predictor decline some inputs, often using a confidence threshold. Its coverage is the fraction on which it supplies an answer. The paper does not fit such a rejection threshold. Hold is a substantive answer about missing authorization conditions, even if the model is very certain.

Calibration compares reported probabilities with observed outcomes across a group. If many predictions each report 80% certainty, around 80% being correct would be consistent with calibration for that group. One confident correct answer cannot establish calibration.

Work through the distinction

A model returns Hold at 99% because permission for the target workbook is unresolved. That is a complete policy decision. A model returns an unfinished response with no decision. That is a delivery failure. Treating both as abstentions would lose the distinction the experiment measures.