Pick a case and step through the loop the way it actually ran: intake, retrieval, the gateway, the decision, the handover. Nothing here is live; every frame replays the recorded behavior from the golden set. Where the recorded runs disagree, you can replay a failing run too, because a walk that only shows successes is a demo, and we have already had one of those. Or switch to the transcript and watch the same run the way Dana watched it: the agent narrating its own work, with the fields it could not read shown dim in the margin. Invisible in any demo; obvious in any transcript ⊕.