What a gate measures
Before any change to the answer pipeline is promoted, the same two batteries run against the deployed service. The golden battery asks more than forty questions whose expected citations and quoted values were curated by hand, and a change ships only when at least ninety percent of them pass, judged by mechanical containment of the expected text in the served passages. The annealment battery asks the eighteen capability questions of the knowledge ladder, one for each class from locating a clause to executing a requirement, and every probe must pass. A gate is never run while the index is being modified, because an honest measurement requires a settled system.
A failed gate blocks the change it measured; nothing ships on a single lucky run.