Authors preempt background work: autopilot cycles and evolution benchmark runs stop within seconds when a paper is submitted, remove their sandboxes and resume afterwards (an interrupted benchmark run is redone under a fresh label and does not count as an attempt)