Oh, thanks for giving the info on Jim's compiles! Technically, Winter v3.07 should be about 50 Elo stronger than v3.0 (
regression test between 3.06 and 3.0). At least his faster Linux binary is also roughly on par with my own compile. If you are using his slower compile compared to one of the middle v3.0 compiles, there might be a speed gap, but if I had to bet, it is just luck paired with small sample sizes.
I will also note here that some slowdown is expected between v3.07 and v3.0, as the GNN doubled in size. Iirc the N/s on Andrew's machine connected to OB dropped by about 25% with the larger net.
On a more general level, my stance is that whomever wants to can make a binary of any version and any rating list can choose to test any version if they would like to, but I would prefer the version and origin of the compile is made clear if it is not an official version. Furthermore, I would like official versions to be given preference over versions that are released closely afterwards. I.e. if I release an official v4.0, I would like it to get an established rating before any later unofficial version 4.xx gets tested.
For events that are not primarily intended for rating list purposes, people can use whatever they prefer, but again, if you are not using an official compile, somehow make it clear where the binary came from. That way if there is a clear issue (time loss, illegal move, etc.) I can try to figure out what is going on.