I don't know of any relevant metrics at all for this sort of thing :-) What we did was mostly review the log of their thinking as they went, then talk to them more to work out if they think like a sysadmin. We want sysadmins who can think, and that's the opposite of a process susceptible to Taylorisation.

Note that in my example, I don't expect them to solve all the problems. I do expect them to have a good try at it, doing it as an open book test with Google to hand. (Open book tests are always hard.)