the main idea behind micro kernels was that they would allow over 10kernels, up to 20, which would mean more harmonics/distortion. the microkernels had been referred to by giancarlo as 'fake' harmonics a few times that i saw. i think they are somehow constructed rather than taken directly from the samples. the reason is because 10 of the original style kernels is pushing what the sampling tech can pull from those samples accurately. the more kernels of harmonic distortion you take from the samples using the math that NAT uses, the more uhh, 'corrupted' the whole set becomes, even including the lower harmonics. in other words, if you create a 3k program (so one kernel for fundamental and 2 for harmonic distortion), those 3 kerns will be more accurate to how they were in the hardware compared to those same resulting kerns if you have NAT create a 10k program. the difference can be subtle and it really boils down to signal to noise ratio. but if you go beyond 10k you get less and less accurate results all around. so micro kernels was like a way of adding 'fake' higher order harmonics above those 10k that are considered 'safe' to take from the actual samples.
i've experimented with 20k programs using the micro kerns and they didn't come out very well. if anyone were to use microkerns for the any of the first 10 kerns, itd be actually a bad thing to do because they are 'fake' and not as good as the full kerns, which are ok to use up to 10k as long as you have a decent signal to noise ratio. you gain nothing by using them and actually are completely sacrificing authenticity in those harmonics. the only reason anyone should ever try to use microkernels is if they are trying to get over 10k for more distortion. this is my understanding after emailing g about it several times and doing a few experiments.