The 4-ply rollout without VR is very similar to the original position:
This should prove that the claim “when the evaluations are clueless since it is just adding noise” is rather unlikely. Especially since one side is almost impossible to play incorrectly, and 84% cannot be achieved by playing blindly.
I’d like to clarify a few things:
I claim that BGBlitz plays certain position types significantly better than XG. These include containment positions like this one (I have several examples that are quite similar) and a post-hit backgame where XG classifies the best move as a 0.300 blunder. By the way, I don’t actively look for such positions; I just come across them occasionally in discussions like this one, and I’ve been writing them down for a few months now.
I’m not claiming that BGBlitz has perfectly mastered outfield primes—it hasn’t—but it understands that the outfield prime has value. What BGBlitz doesn’t yet understand, for example, is that it can safely slot at the top of the prime. But there has to be some work left for TachiAI 7, after all.
The current example is a very extreme position!
What can we expect with perfect play? If the exclusion is successful, about 7.7%. But this doesn’t always work out; every now and then, it goes wrong. How often? I guessed 3%, but Mochy once told me “more than 3%.” He didn’t say how much more, though. Let’s say 10%, meaning we should win 0.9 * 0.077, or about 6.9%
XG is well below that at 0.11% (why is the equity actually -1,000? Shouldn’t it be -0.9978?). Unfortunately, only 24/21 has been rolled out; the other moves are evaluated only with XG++.
Now for BGBlitz. With 1-ply through 5-ply, 20/18, 19/18 is always selected as the move (XG, by the way, never considers this the best move). The rollout (3-ply, VR) yields the following:
For the natural 20/18, 19/18:
For 24/23,20/18:
and for the 24/21 from the XG rollout: 
Far from the expected 6.9%, but 1.5% is significantly better than 0.11%. In 10,000 games, XG would win 11 games according to the rollout, while BGBlitz would win 110 with 24/21 and 150 with 20/18 and 19/18. There’s still plenty of room for improvement, but the numbers are further proof that BGBlitz understands this better than XG.