AI Action Approval Needs Evidence Chain to Prevent Misinterpretation

A developer proposes an 'approval evidence card' to ensure AI actions match human intent, citing risks like a 'update dependencies' approval being misinterpreted as publishing a package. The card includes fields for immutable plan version, actions, destinations, credential scope, reversibility, independent checks, and stop receipt. The proposal aims to create verifiable links between reviewed plans, authority, and executed actions.