The consequence of that is that it has to be reliable enough to play through a full concert without maintenance or breakdown, and it has to be robust enough that it can be transported from place to place. These are his hard requirements.
Then there are some less well defined requirements. Which is that the machine has to play nice music and has to be a marble machine as Martin understands it.
This last is the real constraint. Otherwise he could just buy a midi keyboard which would fulfil all the requirements about reliability, robustness and quality of music, but would fail the spirit of the endeavour.
I agree that there Martin seems to be aiming for a very high degree of repeatability in timing, but it also seems that he has designs which meet those expectations of his and this was not the reason why he abandoned the second attempt. (Ad far as i can tell based on the videos.)
Many of the digital sequencing and notation products I've worked with went out of their way (arguably) to play "less-tight music" through various "humanizing" features.
Yes, we want music that is sufficiently accurate and "tight"... but within the confines of human capability. The slight errors of both time and intonation in some cases give music a much more human feel. Now to be fair, I don't want to suggest that this sort of human inaccuracy is mere randomness either: it's typically not just random error... there's usually a bias and it definitely within limits (unless you're a bad musician of course :-) ).
Super-tight MIDI music sticks out like a sore thumb when you mix it with things performed by humans.
https://www.youtube.com/watch?v=WN90HYiFpAw
Some people will say it's common-sense stuff but it is stuff I see everyday writing software and it's so hard to change. It's refreshing to see a spaceship company having the same issues haha.