Fascinating stuff. I fully understand the appeal of off-loading the bias control/protection functions to software, with all the clever nonlinear stuff it could do, but I can't quite work out what this implementation is doing that results in it going into protection mode if a continuous tone is playing and why that could ever be considered acceptable. For a start, such a system would preclude the amplifier from any use other than 'domestic audio' which seems like a peculiar limitation to place on it. Woe betide anyone who wants to set up their reel-to-reel with test tones or whatever.
Is it that they are just using a low powered microcontroller and 'sparsely' sampling the amplifier output (they mention "30 times a second")? Statistically it will work OK with most 'music' signals, giving an output that averages to zero. However, it will run into trouble with a fixed tone, with regular aliasing possibly resulting in a non-zero DC indication - which their algorithm interprets as "runaway bias".
If that is the case, the solution would be to sample the amplifier output at much higher sample rates (e.g. 192 kHz) and perform full-blooded DSP on it. But isn't a hybrid DSP/analogue amplifier a bit 'ugly', anyway? Such a beast seems a bit silly when put next to a fully Class D amplifier.