AI Approvals Should Expire When Underlying Evidence Changes, Not Sessions

A software design proposal argues that AI-generated action approvals should be tied to specific data snapshots rather than browser sessions, expiring automatically when material facts change. Under the proposed model, an approval authorizes a versioned 'action envelope' that specifies exact scope, consequence, and evidence revision, becoming void if any critical field is updated. For example, if a reviewer approves 12 refunds totaling $430 but the dataset grows to 19 orders before execution, the system should block the action and present a clear diff of what changed. The author outlines rules distinguishing material changes that must trigger re-approval, such as shifts in affected count or total cost, from minor wording updates that should not interrupt the reviewer. The proposal also calls for server-side envelope comparison, accessible change notifications, and scenario-based usability testing to validate whether reviewers can accurately identify authorized scope after a change.