Extracting a conclusion from the froth of a sea of arguments.
If we could pick a random sample of arguments, then we don’t have to evaluate their entire population to reach a reliable conclusion.
But instead we are biased, so we don’t trust anyone to pick a random sample. We just give in and let each side pick only arguments in its favor. So we search for arguments incrementally on both sides, and only when one side hits diminishing marginal returns sooner and more drastically, that it becomes apparent that the arguments overall support the other side.
Furthermore, this process is made to be as drawn out as possible, since each side does not want to mine their arguments so efficiently that they would hit diminishing marginal returns sooner than is necessary. All they want is to find arguments strong enough to immediately counter those on the other side.
This is analogous to signaling, where the bulk of the total amount of resources spent on signaling does not contribute to discriminating between the signalers. To make yet another analogy, imagine a 100m race. What determines who finishes the race in which positions is their relative positions when they cross the finishing line. The finishing line could have been placed at the 50m point instead, and it will still be as discriminatory. Now of course, we don’t place the finishing line at 10m, because the longer the distance the runners run, the more random deviations from their average speed evens out, and we are really interested in the runners’ ability, not their luck. Arguably though, signaling contests go far beyond what is necessary to float the signal up from the noise. That’s the contention of this post. Too small a random sample of arguments leads to vulnerability to noise, but we take too large a sample, because we cannot trust it to be random.
Let me end with a favorite metaphor of my own.
Signaling is a spitting contest in an arena of froth on the surface of a sea of saliva.
Having read this post, you should be able to get what it means. Otherwise let me know, and I’ll do a post on it. Or maybe I’ll do a post on it anyway.
sarkology 
I am going to use your 100m race analogy.
Signalling is not over until the finishing line set by norms is crossed. Norms often don’t set finishing lines, they set noise floor. Relative deviation at 50m is barely above the noise floor(which is a noise level below which signalling isn’t effective). Since the runners don’t know where the finishing line is, they keep going, bulk of their resources from that point onward will be spent on signalling maintainance.
The problem, I think, is that there is no halting command. In reality, 100m race runners would only slow down and halt their signalling when they crossed the finishing line that has been set by norms, otherwise they would just keep on going and the relative deviation gap and individual signal-to-noise ratio would increase over time. Thus, resources can be saved by setting a halting command, in this case, the finishing line of the 100m race.
So, the reason why extracting conclusion could be inefficient is because there is no set ‘finishing line’. The results that were produced by confirmation bias and conclusion that can thus be drawn from, could all be done at the 50m line, only if 50m was the set finishing line. We can see that confirmation bias(which is analogous to each of the runners running in their own lane) that run amok up to 50m can produce confident conclusion, which in the case of 100m race is the anticipated winner of the race.
I don’t think social norms need to come in to set the noise floor or the finishing line.
The noise floor is naturally set by whatever random influences affect the achievement of the competitors.
The finishing line is set by competitors giving up.
But you are saying that IF social norms set the finishing line, then we can save resources. That’s true I guess, so maybe like in a debate, the competitors have a certain amount of time to come up with arguments, after which we just decide based on what they managed to come up with.
So how can we get such social norms?
I don’t think you’re taking into account the inherent reward for signaling.
Can you elaborate? Not sure what you mean by “the inherent reward for signaling”.
The 100m race analogy would be very close if it was a race where the contestants could call for the finish line to be pushed forward at any time during the race. Eventually they would give up when it was clear that they had faster diminishing returns to running.
Neat!
We don’t let such runners get away with it, yet we don’t do the same for people making arguments.