I'm probably saying stuff you've already guessed, but for completeness... It's less about the benchmark per se and more about having the WR on a benchmark.
While doing bunch of lower clocked runs, they considered what they might WR and looked at the hwbot.org charts (if they didn't already know the charts). The system was likely not fully stable, so perhaps the kind of instability the system had at that frequency, the GB3 workload (https://www.geekbench.com/doc/geekbench3-workloads.pdf) is less susceptible too. Off the cuff, it is over a decade old and probably hammers SIMD instructions like AVX* less (if at all).
