Test whether safety framing changes agent performance with matched tasks, paired metrics, refusal transparency, and monitor-recall gates.