I think one cannot generalize on the 7->10V stage based on a single (or couple) of successful builds with such small TCs as above, imho. Those are almost outliers
What Dieter indicates is that the version with the ovenized 7->10V stage is a safe and least expensive solution than an elaboration with statistical arrays (randomness) or manual fine-tuning of TC (time, labor costs). It means ie. with an 0.1C oven and 5ppm/C resistors you will always get the total TC<1ppm/C with no effort or risk.
Well, this technique of T.C. matching of resistors and T.C. characterization / trimming of the complete circuit has been done by FLUKE for all of their old reference instruments, like 730, 731, 332/335, and this statistical array plus T.C. trimming has been done by John R. Pickering in his Datron / Fluke 7000 references, very successfully.
I don't regard an ovenized 7=> 10V stage neither as a least expensive, nor as a simple solution, especially in regards of volt-nuts, who mostly strive for most simple and cost effective solutions, which also work practically. That's always my approach, to invest some good ideas and a bit of measurements for simple solutions, instead of over-engineering things. Btw., I work in R&D of the Automotive Electronics industry, where you have to save every tenth of a penny.
I can't remember that anybody has published his fully ovenized, practically working and fully characterized (10V) LTZ reference here. So, please think again about "no effort, no risk".
It's clear that my low T.C. 10V reference is achieved by chance, but I never made any statement, how probable or improbable this is. I also did not state any generalization of my result.
Therefore, I don't value both your statements, either that the likelihood for a pure resistive divider is rather high to end up with > 1ppm/K, or this is almost (always) an outlier result. None of us has a statistical idea, or can present a test series, how the probability for a good T.C. matching is, for the different suppliers of such PWW or BMF resistors.
Anyhow, John Pickering used 2 EA of those TiN (?) resistor arrays, in an entangled schematic, to create a stable statistical divider. It's obviously very successful in series production, as they could guarantee a very good timely stability .. the T.C. is trimmed internally, but I assume, that the overall T.C. is already quite low as determined by the arrays themselves.
branadic made similar experiments with a single array, and trimmed the T.C. by means of a short piece of copper wire. The timely variation was quite high, afaik.
If I understood Andreas correctly, he achieved a very low T.C. with a single (?) array as well.
Therefore, this method is not that elaborated, but already approved for volt-nuts.
In the end, why I insist on simple, already validated solutions is the following:
I know Andreas lab on-site, and RAX also described his man's cave in detail. Both labs suffer from big temperature variations (+15°C / - 5°C, or so), therefore, to anyhow do ppm/sub ppm metrology, one needs ultra low T.C.s for the references. I mentioned my exchange with AB Precision @ 0.3ppm transfer level; both our basement labs might have different room temperatures, (in fact, exactly the same), therefore this requirement is as well mandatory.
Even with such very constant temperature conditions (+/-0.2°C) in my lab, the residual T.C. of any precision instrument can be quite annoying .. like the T.C. of a transfer instrument like the HP3458A. I promise to soon write an article about my recent T.C. measurements on the 03458-66509 reference board, and the overall T.C., and how this might disturb sub ppm transfers.
Frank