That’s a very odd result indeed. Samples that receive an accuracy score belong to the so-called gold standard (references) group.
After watching ccp stream on Project Discovery, where the chief scientist involved (Ryan Brinkman) explains a few things in more detail, I would definitely “gate” four clusters (3 in the lower half, 1 in the top half). All those clusters would be populated significantly (have centers colored differently than just the low population blue code). That is exactly what the scientist on the stream said is the next step for better quality analysis. Moreover, both axes (horizontal and vertical) in these 2D plots are logarithmic and cover three orders of magnitude each…
I guess this is an old “gold standard” from the very beginning of Project Discovery. It should be replaced or deleted, because it actually forces you to make a wrong analysis and the wrong conclusions.
P.S. Dr. Brinkman did mention that consistently bad gating by contributors leads to their analyses being taken out of the study (as I understood it, part of this project involves machine learning)