I don't believe in matrix-only profiles for general-purpose work due to 1) how do you apply a contrast curve to a matrix-only profile? I don't think RGB curve or Adobe hue-stabilized version is the answer, and a linear curve is really only for repro work 2) raw converters expect the profile to deal with gamut compression, and you can't do that in a matrix.
That said I do believe that a LUT working on the colorimetric level should not make large changes to the matrix, and you can see that in commercial profiles to that they have similar approach. The LUT contributions is mostly about tone reproduction, ie the curve, and gamut compression.
The GUI version of DCamProf will allow manual matrix tuning, but not really by entering numbers manually, but through manually-controlled refinements on the automatically derived base matrix, it's basically a graphical version of the make-profile/-v parameter of the command line program. I've thought about including numbers/sliders though for those that would perfer that type of direct tuning, probably not in the first release though.
DCamProf matrix optimizer is currently limited to make white-point preserving matrices, that is you cannot make a matrix that tints the neutral axis (you can make an ICC profile that does that, but that's about white balance compensation which is another thing). I have so far not seen any reason to change that. DNG profiles actually requires the ForwardMatrix to be white-point-preserving or else Adobe Camera Raw won't accept the profile.
I often do tweak the matrix by hand, "because I can". My favorite modifications is to match skin better and let others color suffer a bit, make sure deep blue is not too dark and is rather turning towards cyan than magenta, and also looking at the reds which I rather have warm than cool. Then the LUT make minor adjustments on top of that. Greens are generally stable as they have strong contribution from all three channels. This is for daylight. For StdA I still need to gain some more experience.
Lightness control is an issue for matrices. Matrices have pretty large lightness errors in general, but I think it's not a big issue as while we can detect lightness errors very easily in an A/B swap, we don't really remember lightness well, or maybe more accurately put, a color slightly light or dark is not seen as unnatural while a hue that is off or under/over-saturation can be quite conspicuous. I think this is because in an image lightness could be the result of the light in the scene, so the eye doesn't really know what's the right level, while (normal) light doesn't cause hues or saturation to change so we react to that even when we have no reference to compare to. So I think lightness should be the least thing to focus at.
To make the matrix "mallable" it should not contain any strong negative factors, which can lead to clipping. One thing I've worked with quite much in DcamProf project is the overly sensitive blue channel (probably for better ISO and tungsten performance) which cause strong blue subtraction in daylight which can cause clip to black for saturated colors. The solution is to sacrifice lightness accuracy and lighten deep blues significantly. Likewise early results suggest that reds must be lighter than natural for an StdA matrix in order to provide stability.
