If you attempted to solve multiple times on a randomized cube you are eventually going to stumble across some faster solves, how do they remove the luck element?
A fair question, and one that I've thought about. The final judging attempt hasn't happened yet, so it will be somewhat up to the judge.
It will definitely be a hand scramble by the judge. A machine scramble would open the door to play back a memorized reversal.
There are websites ( like cubtetimer.com ) that will generate instructions for a "good" scramble, but there is no guarantee.
We don't try to generate a fully optimal solve ( takes too long, and we are counting solve time as well as move time ), and we get variations from 0.9 seconds up to 1.4 seconds. The even the slow end of the range is enough to claim the current record.
The 20 move longest solve is measured in the "quarter turn metric". Our machine can do half turns, and can combine opposite face turns, so our command list isn't directly comparable to the quarter turn metric.
kociemba.org is a fascinating website. They show a histogram that shows that about 70% of solves can be done in 18 moves ( in the quarter turn metric ). Scrambles that require a full 20 moves are exceedingly rare. You are not likely to witness one from a random scramble in your lifetime. But they are pretty easy to engineer.
Also ( added with an edit ):
From the WCA scramble rules:
Puzzles must be scrambled using computer-generated random scramble sequences.
Generated scramble sequences must not be inspected before the competition, and must not be filtered or selected in any way by the WCA Delegate.
So as long as the scramble is random, the don't appear to enforce that it is hard. But the odds of getting a really easy scramble ( solvable in less that 13 quarter-turns ) at random appears to be less than 1 in 10,000
The sweet spot for random scrambles is the 16 to 19 quarter-turn range, with 18 being the most common.