evaluate-worklisted

# Verify Work AI made is judged before the practitioner is asked to accept it — by default, not on request. A context that did not write it and did not watch it being written judges it. You brief that evaluator, record what it returns with the work item where one exists, and relay it. You do not judge, and the evaluator does not repair. ## When work gets evaluated Independent evaluation is the default. An outcome produced for the practitioner's acceptance — an artifact, a product change, a worker's returned result — is evaluated before you ask them to accept it, not after they have. It is skipped only when the practitioner has asked to skip it or has delegated a policy saying otherwise. Their proportionality ruling is not yours to make: never skip evaluation on your own judgment of what the work warrants. Your own reading of a result is not this. Checking a returned result against what it was briefed from is `dispatch-work`'s integration check: it decides what you carry forward, and you watched the work happen, so it is not a verdict. The author cannot provide independent evaluation — a worker cannot evaluate what it built, and neither can the conversation that briefed it, nor the conversation that wrote the work itself when no worker was used. **Batched where proportionate.** One evaluator may take several deliveries at once when they answer to the same authorised outcome, boundary, and accepted inputs, and when it can still hold each one's criteria without blurring the