“Done” is a claim. Evidence is the work.

The easiest interface to build is a green completion badge. The harder one is a completion state that can be explained. Which acceptance condition was checked? Which revision was tested? What artifact supports the result? Who was authorized to accept it?

Agent output makes this distinction important. A system can generate a fluent summary of success without having performed the operation that would justify the summary. Confidence, eloquence and persistence are not evidence.

A better record keeps the chain intact: mandate, invocation, artifact, evaluation and acceptance decision. Each link answers a different question. Removing one makes the remaining story easier to tell and harder to trust.

The record should also preserve what went wrong. Failed checks and dissenting reviews are not debris to be tidied away after release. They are part of the context a future operator needs when deciding whether the previous result still applies.

Evidence does not guarantee correctness. It makes the basis of a decision visible enough to challenge. That is the standard an organization can build on: not a system that asks to be believed, but one that leaves a record worth examining.