Does YC not run a risk here of falling into a reinforcement loop where each generation gets more and more "perfect" from the POV of the current generation, leading to highly specialised but unadaptable individuals a few generations down the line?
S-curves usually form in situations where the people doing the selections are trying to impress a third party, e.g. US News & World Report, future dog purchasers, other peahens, etc. Fortunately that shouldn't be a problem here.
Also, because YC isn't very bureaucratic it should be relatively easly to design a process that ameliorates this. It's an interesting problem to think about though.
Adding some randomness could help to avoid over-fitted populations. Then you could prevent to converge to a sub-optimal state and force the system to look beyond of what it seems the best.
