It's probably like venture capital. There are many scientists who test many hypotheses. Many are bad at generating hypotheses or running tests. Some are good at one or the other. Some are good at both and just happen to pick the ones that don't work. Some are good at all.

But you can't tell ahead of time which one is which. Maybe you can shift the distribution but often your pathological cases excluded are precisely the ones you wanted to not exclude (your Karikos get Suhadolniked). So you need to have them all work. It's just an inherent property of the problem.

Like searching an unsorted n list for a number. You kind of need to test all the numbers till you find yours. The search cost is just the cost. You can't uncost it by just picking the right index. That's not a meaningful statement.