What problem does it solve? AI agents fabricate citations, overstep authorization, and report "done" without proof, leaving teams unable to trust delegated work. This Skill replaces blind trust with executable verification: acceptance checks, judge layers, and audit loops that catch failures before they cause damage. ## Core Features & Use Cases - Delegability screening and constitutions: Sort workflows by whether an executable proof of done exists, then write 10-14 testable acceptance criteria per delegable workflow. - Judge layer design: Build actor/judge separation with allow-block-revise-escalate outcomes, a four-class action taxonomy (read-only through high-risk), and structured action-proposal schemas. - Auditions and monitoring: Run scored tryout tasks before granting autonomy, then track completion-vs-acceptance metrics and feed user corrections back into judge eval suites. - Use Case: Before letting an agent send external emails, classify the action as class 3, wire a judge with authorization and evidence criteria, and require a structured proposal stating recipient, source, and rollback path before execution. ## Quick Start Audit my agent workflow for trust gaps and design a judge layer with acceptance criteria for its highest-risk actions.
