I finally got around to a few tests myself. I had some MPF102 (plastic) so I decided to test them. The zero temp coefficient point was only 1.2V at 1mA, but this can be useful if you need a reference that can run from 5V.
The voltage sensitivity with a 5V supply was 0.003V/Vsupply. At 15V it reduced to 0.0013V/Vsupply. I am pretty happy with both numbers. As long as you maintain the voltage across the reference fet to 100mV, the voltage sensitivity even with a 5V supply is negligible.
Resistance sensivity is worse. For both 5V supply and 15V supply, the output varied by about 1/3 the resistor variation. This does not include the fact that as the resistor drifts, the FET current drifts and so it gets a temperature coefficient. The TO92 package is not good for temperature tests as you can get a positive or negative change depending on where the heat is applied to the package.
There are a number of 3ppm to 5ppm reference ICs for around $3. To equal this, I will need 10-15ppm resistors.
So on this score, it may not be worth the effort.
The other issue is the long term stability and temperature hysteresis. For reference ICs, it is not unusual to see a 50ppm drift spec for 1st 1000 hour drift and another 50ppm drift if the chip is temperature cycled from 25degC to the min and max temperatures and back to 25 deg C. It would be interesting to know what kind of drifts and temp hysteresis you get with the JVR reference. It is certainly cheaper and easier getting a hermetically sealed JFET then getting a hermetically sealed voltage reference IC, but of course, the resistor needed for the JVR is probably not going to be hermetically sealed.
If it turns out the JVR is low drift/low hysteresis, it is interesting for custom jobs, especially if the JVR is put in an oven.
Anyone have some suggestions for economical and stable resistors in the 5ppm to 15ppm range?