The idea: the critic gradient tells each action dimension which way to improve -- but it's only reliable within its support. Replay actions are always in it (the critic is trained on them), so QF3 clips the flow's velocity to stay near one, "filtering" the Q-gradient past it