Comment by svnt

It is the benchmark error rate, not the benchmark success %, that we actually trip up on.

Going from 85% to 90% is possibly 1/3 fewer errors or even higher, depending on the distribution of work you’re doing.