Going below a quorum of 3 might produce incorrect results on some occasions. We run a huge amount of work and once in a while we see a duplicate failure.

However, we are going to try a new backend validation approach that will effectively need almost no duplication of work, yet yield quality results. This will only be possible for certain kinds of research projects, such as HPF2, which should resume soon. For those kinds of projects which have many calendar years of work, this will be a substantial speedup and a great benefit to the researchers. We will see how it goes with HPF2 and if all works well, certain other of our projects will be able to take advantage of this approach as well. So please stay tuned.

That's exellent news indeed.

Just another thought (slightly off-topic):
You state that there will be projects which last for many calendar years. In that case it might be worthwile to develop specific clients for specific instruction sets, and even maybe for GPU's (IF they prove to be as fast as promised).
Of course I understand that for a relatively short running project there will be not such clients on short therm, but for projects that run for a very long time, it might be very interesting.