The selection process seems to do a good job of keeping out the lowest tier at least, although we openly acknowledge that we miss a lot of good people as well.
Years ago, I worked someplace where a colleague was tasked with working with another developer on project X. After about 15 minutes it was clear the other developer ... wasn't? A web project, and this person had been employed as a "web developer" for at least several months. Questions like "how does this information in this browser get back to the server?" came up.
Colleague goes to manager and says "I can hit the project deadline, or I can make sure other_dev learns the basics enough to be able to contribute and understand projectX, but I can't do both by the deadline. Can we move the deadline back a few weeks?"
No, and no. Train other_dev and hit deadline.
Deadline was hit, other_dev moved to another project afterwards, and was pretty much as ineffective as before, but colleague was then saddled with this reputation of being a 'bad mentor' because the next team learned other_dev didn't know how things worked. Why the hiring manager wasn't tarnished... who knows?
He was given 2 tasks and he only delivers one result.
I know this sounds... not ideal... but it is what it is.
His manager probably has to operate at the same level of expectation: given 3 tasks by his manager (or director), either you finish all 3 or you're less dependable.
On the other hand, if you think the software industry has a hard time figuring out (at hiring time) who the high performers are... science is driven by serendipity. Nobody can predict who will find the billion dollar discovery. Not even past performance is a reliable indicator.
So it makes sense to me that the salary spread in science is relatively even. If they could reliably figure out who to dump money on, they would. On the other hand, the FAANG companies clearly believe their hiring practices can select out the high performers... and perhaps they are right? If they're paying 3-4X what everyone else does, they expect to get at least 3-4X the value.