A property of a control architecture in which AI-initiated actions can be reversed to a prior state, characterised by the availability and coverage of undo or restore mechanisms across the system's actions; an affordance rather than a discrete event.
STATE. Operationalized as reversibility coverage: the proportion of the system's consequential action types for which a documented undo/restore path exists, scored 0-1 from the system specification. Optionally banded (none / partial / full). It indexes designed affordance, not how often undo is used.
Proposed measurement protocol (not yet empirically validated): Specification audit enumerating consequential action types and checking each for a documented reversal path; reversibility coverage could be reported as the proportion reversible. Two auditors independently classify a sample; agreement on the per-action reversible/irreversible decision could be reported as Cohen's kappa.
https://andreasehstandlicenseofclarityloc.github.io/augmanitai-periodic/#topic-agentic-ai