I spent some time yesterday looking at some of TiN's data. I've very grateful to him for making those available. Otherwise I should have nothing with which to work. I cannot justify expending the amount of money he has on lab equipment and I have no way to arrange access to such equipment. I just bought an HP 8560A w/ TG and eagerly await its arrival. And I just finished cleaning my 3478A. So my lab is quite modest.
One striking aspect of the data is the amount of noise present in the measurements using a 3458A. It is my understanding that the instrument will average measurements over a user selected period, commonly a number of power line cycles. Thus I expected far less noise in the data.
In looking at the 3458A operator's manual, the discussion of the setting of bandwidth and integration time is rather vague about the window edges. In particular, if one integrates all samples over a number of power line cycles without applying a weighting function to the first and last few samples, it corresponds to a rectangular window in the time domain which is a sinc function with significant sidelobes in the frequency domain. This allows a considerable amount of noise leakage, far more than one might expect. Moreover, such noise will be aliased quite severely.
Does anyone know of professional papers on the topic of the analysis and reduction of noise in high resolution DC voltage measurements? While a simplistic treatment in an operator's manual would not surprise me, I should be quite surprised if the actual implementation is not more sophisticated.
White noise (above 1Hz) can be filtered out, LF ("1/f") noise (aka: "pink noise") is extremely difficult to filter out. There is no lower frequency limit-- as LF noise can have periods that exceed 1 month or even more. So, the total system noise is the RSS of the noise of the device under test, and the measurement instrumentation noise-- and strangely, longer integration times can actually spoil the measurements. For any given measurement system and device under test, there is an optimum integration time that can only be found by experiment, and then the Allan variance can be plotted-- which will indicate where the "sweet spot" is for integration time.
Sometimes a simple gathering of a small (and odd) number of measurements, followed by a median filter will yield the best results.
The BEST way to filter out 1/f noise is to not generate it in the first place. So, you select a reference source and device under test that has the smallest amount of 1/f noise that you can find (and/or afford), and go with that. You have to be realistic and understand that unless you have quantum based intrinsic standards, you are going to have to accept some level of noise in your measurements, and the final uncertainty will not be zero.
That said, as an example, you can put a 100mHz filter on the output of a voltage reference like an LM399, and this will reduce the apparent noise by about 6X (1uVpp)-- but you will NOT be able to filter out changes in the LM399 that happen over minutes (about 1ppm jumps in the reference voltage). So, because of this, the LM399 cannot be relied upon for anything better than about 1ppm. If left on continuously, the LM399 will exhibit a temporal drift (of typically 4ppm/a, but some are better than that). If you only turn it on for a very short period to take a measurement, then turn it off for a very long period, the long term temporal drift can be negligible, but you are still limited to the 1ppm uncertainty due to the 1ppm "jumps" in the output (due to low Zener current, which you can do nothing about). If you are looking for better than 1ppm uncertainty, then an LTZ1000 is the only commercially available reference that has a long track record of providing sub-ppm uncertainties and annual drift rates. (Obviously, some LM399s and LTZ1000s are better than others, and typically 1 out of 100 will be exceptionally good).
The same goes for resistors-- there are many known drift mechanisms and ways of dealing with those drifts. A hermetic package eliminate a whole set of problems. After that, artificial aging can significantly reduce temporal aging effects. After that, placing the resistor in a thermally lagged enclosure can improve temporal drift, and short term temperature drift. Keeping the resistor in a thermally stable environment (oil or oil bath) can further reduce temporal drift. Different resistance wire materials can have very good temporal stability-- some better than others. Zeranin-30 and Evanohm alloys are probably the most stable.
So, this was a very long and complex answer to a simple question about a very complex problem. The short answer is that you CAN help some with sophisticated filtering, but at the end of the day there is a limit to what you can do with this, and I think you will find that your money and labor would be much better spent on better standards and measuring tools.