The Valhalla 2720GS (Vdc calibrator) is doing something like that. There is up to 4 boards with 2 vref per boards. I'm not sure if the vref on 1 board are averaged or if they are tracked individually.
some info on the 2720GS: https://doc.xdevs.com/doc/Valhalla/2720GS/
Ah, prior art! The datasheet says that they individually digitize each output using a 30-bit ADC. I didn't find a text description of the circuits (only found schematics), so I don't yet know if it is doing single-ended or differential measurements between the reference and the DAC.
I hadn't considered batteries, but that does make sense as a voltage source. Personally, I was thinking of summing the LTC6655 output with a DAC output to get 10V. The LTC6655 would have the advantage over the battery of a much lower TC. I don't know how the noise level would compare between the LTC6655 (or other references) and a battery. However, with more transistors/die area in signal path of the LTC6655, perhaps the chance of RTN (popcorn noise) is increased.
What is useful to me would be to build a low voltage calibrator (10V, 1V, 0.1V) to help me synchronize my DMMs. I don't personally NEED a ultra-low-noise 10V reference, but it'd be fun to make.
There is still the trouble of going from 6.2V to 10V. It needs a highly linear ADC. Doing the differential measurements as I proposed only helps with outputting a very low noise value close to the "raw" reference voltage.
(I find myself writing too many words; I hope that they are at least a little bit interesting. I will start my own thread if I continue with the above idea).