This is, I fully admit, the single most confusing thing, and something that always comes up, so I'll break it down as to why it looks that way. All of these patterns have to be encoded at YCbCr 4:2:0 since they are on a Blu-ray disc. In order to get a pattern that has, for example, Red 234, it is going to have corresponding blue and green values as well. These values for the test pattern are always under 16, which for most video cases renders them as virtually insignificant. However, for Red 234, our RGB reference value might be (234, 9, 11), and the value being returned is (234, 10, 11). In that case, there is no error for the Red value, which is what we are looking for, but there is a very, very minute error in green, which causes a dE error still.
I've been working on ways to better process and evaluate this data, and that might include running dE tests where we consider those secondary values, and ones where we don't, and then seeing what the difference is. Since those errors cause only around a dE of 0.2, which is so insignificant as to make it almost meaningless, I wouldn't worry about them myself. That said, some players do get those 100% perfect, and I think we should recognize those players. Some take any value below 16 and make it 16, and while in most cases that will be fine, there is no reason to do that and so recognizing players that don't do that is important.
I hope this clears that up. I could put all the data on the results page, but then it becomes incredibly messy and hard to deal with, and I don't think that makes it better for anyone if they can't read the data. I could also just ignore those numbers, but the last thing I want is someone to accuse me of making the numbers look better by playing around with them instead of being totally accurate.
Did you test consider dithering (if any)? If a sufficiently large flat patch was used, spatial averaging of the output should return the "pre-dithering/quantization" value that we probably are most interested in.
It would also be interesting to see if >8bpc HDMI links actually contain useful information in the lsb (given that the source does any processing that warrants more than the 8 bits of Bluray), or if they are just zero-padded.
I haven't really tried the DeepColor testing yet, where something like a color gradient would be more useful than color patches for that to see how it interpolates the values between two figures, but I have thought about the dithering issue. Right now we know that some players dither, and so reference values can be +/- 1 on some players. If I think I notice this going on, I usually check out more points and see if it really is dithering it, as we don't consider dithering to be bad.
I'm starting to advance the test processing to the point that I might be able to sample 9 or 25 points instead of a single point and then get an average and median value from those, which would put to rest any worries about dithering in a sample. That might be the nest target I am for on the test, as it leads to even better results, which is what we want.