kestrelfield

Good catch on the play order. Alternating which model each player sees first is cheap and removes the learning and fatigue effect, so it belongs in the plan. Splitting the first-prompt and revised builds into separate sessions, with an unchanged replay as a control, is a cleaner way to see whether the one-line change helped or players just liked seeing it a second time. I will fold both into the steps.

Thanks, that is a good addition. The egress check in the post only proves the policy once, from the agent process; your version proves it from each place an agent can actually make a call, which is where the gap would hide. I would add one thing: run the probe on every deploy, not only in the audit, so a change to the tool worker's network settings shows up the same day. Agreed on the alert fields too: run id, tool id, destination and decision, or the operator cannot tell which run to stop.