@ MisterDiodes: Always a pleasure to read about your experiences, especially since i have the 2508-32-Evalboard sitting in a corner!
Maybe you can have a look at the proposed voltage-reference-filter-design mentioned here and share your opinion: http://electronicdesign.com/energy/filter-trims-ultra-precision-voltage-reference ?
Also: do you know any good ref-filter-design for a LTZ1000 regardless of the stabilizing-time?
Edit: Since the LTZ1000/Ref-Amps are the best voltage references out there apart from the superior JJA and companies like Fluke/Keithley surely like to improve their stuff: Do you know of any successful projects which improve stability/noise of those references or does it just not make sense and the only good way is to handselect and parallel those references like it is done in the Datron 4910?
Ha! That circuit's been floating around a while for sure, in various forms. Nothing to see there, and a poor ad for Maxim - another capacitance multiplier. A great teaching moment in 1/f noise and how hard it is to get rid of.
Just try building a LP filter with cutoff down in the 10's of mHz - watch what happens when you turn it on. You'll be waiting a while... <Grin> You'll still have noise over longer time spans now, and you'll be kicking yourself for blowing that much on those $65 caps, etc.
You don't need the "Teflon standoff's" either if you understand guard circuits - and you sure don't need that extra shunt regulator. That was there just because those goofy Maxims were only 5V parts - typical of the early crop of AZ amps. Noisy buggers also, compared to later offerings. Maxim seemed to always miss the mark there.
Not making noise in the first place is the concept, and paralleling Vref's (up to a point) is probably the better bet. Probably 9 or 10 Vrefs is a practical limit and point of real diminishing returns. 4 Vrefs might be even a better upper limit. Only if the customer is willing to pay though.
If you build this stuff for a living, you'd better make sure there's a practical sense of profit also. If you chase low PPM's for no real reason and no profit - you'll be out of business.
At some point you have to tackle the problem the other way and decide what noise you can live with...and what is "good enough" if you have multiple measuring systems. The 3458a has great linearity and has good resolution for relative measurements over shorter time spans. As good as it is you still have 1hr, 10hr and 24hr uncertainties that are unavoidable - and that's why the most practical approach is to average across multiple systems
(Vref and ADC or whole DMM's) to help deal with 1/f noise.
Better yet is to spread your measurements across different devices and techniques
- so that why we have 3458a's as well as 732's, KVD's and null meter sets. That way you get a better spread of drift rates and noise, and more likely to fence in an accurate estimate of a DC voltage. What you don't want is everything drifting together with similar 1/f noise characteristics.
Remember: It's always only an estimate measure when you're in low PPM territory. Even a JJ-Array has uncertainty. Most of the time trying to spend an exponential amount of effort trying to shave another half-ppm of uncertainty is a major waste of time AND money.