I think a very basic reason it shouldn't be possible is that for really accurate sync each device has an unkown bias (innacuracy) besides just high std dev (imp

I think a very basic reason it shouldn't be possible is that for really accurate sync each device has an unkown bias (innacuracy) besides just high std dev (imprecision) with relation to it's network/processing behavior. If you average out a lot of samples you can get rid of all imprecision (albeit very slowly), but you essentially need a very accurate clock to find the bias, so it makes the whole network approach useless. I would guess you couldn't even solve for bias with a clock once and then keep using the network, because the bias should be time-varying.