Corrected some of the measurements to fit to the second INL run, which makes it even more obvious that up to now nothing has changed at all. Tested also a few other things, such as 470pF directly connected without additional resistor, as I didn't wanted to cut traces by now.
I also tried changing the integration time, with respect to the two papers "Precise Measurement of the Accuracy of 24 bit ADC by AC Josephson effect" and "Application of the AC Josephson-Effect for Precise Measurement" 50NPLC and 10 samples are mentioned: "... the offset error of the ADC is measured as a difference between the reading value of the ADC (average of 10 data obtained with each SO PLC (1 sec.) integrating time and same auto zero adjusting time) ...".
But that didn't change anything either. I have now installed a 5.6k resistor to U206 PIN6 to PIN4 and have the measurement running back on 20NPLC. Unfortunately it needs the full run, as during the first run the meter warms up, while the second run is used for analysis or at least, it needs the time for everything to thermally stabilize anyway.
What we can conclude by now is, the design is fairly stable, even with this INL issue.
But I'm not yet about to give up.
-branadic-