I'm not seeing good evidence that journals or conferences do any better of a job at this than pre-print servers do. They do have a slight edge, but that is not that great. The edge is because the default position is to reject and so identifying just a small percentage of good papers does affect this outcome, but that's not a meaningful signal. Have an extremely high false positive reject rate and a high false positive accept rate doesn't make the signal meaningful, it just means it is very noisy.

I don't give this very low signal any meaning because the truth is that good works filter to the top regardless of the journals. They rise because peers share them, not because they got the stamp mark from a journal.