The scores were selected based on the expected best-case-scenario score, which is the WR for each platform (so Haswell). From that WR, we took something like 70% to set the baseline.
It used to be round numbers, and that is arbitrary. By using a reference score to get the baseline of, there is some logic behind it. If you take arbitrary numbers, there is a benchmark scaling bias. By this I mean: with arbitrary numbers, you don't take into accoun the scaling characteristics of the benchmark. You end up with the situation where the SuperPI-32M score has almost no effect in the final score.