@Kleinstein: I fully agree that the 2N4416A is not the best choise. I will assemble references with different JFET types. As a consequence of my bad experience with humidity diffusing into plastic packages, I prefer the TO-18 case. By the way, I observed random telegraph noise of the 2N4416 which is most intensive at a particular temperature, while there are temperatures free of RTN. The JFETs don't have a stress reducing layer below the chip, as it can be seen at some references. It seems that you can't get JFETs in an optimum case.
Possibly someone is interested on my reasons for testing JFET references:
Zener diode references base on a controlled avalanche breakdown. My understanding of avalanche break-down is as follows: after ignition of a microplasma, its current drain is limited by the doping concentrations in the n and p layers (roughly a resistive feedback, the current density is limited). There is a competition of microplasmas: the microplasmas with the lowest break-down voltages determine the reference voltage, since the total diode current is well controlled. Richie made excellent pictures of this avalanche break-down in the LTZ1000, where a limited number of bright points is visible in the depletion zone. Now, microplasmas are switched on at any imperfections of the Si crystal. A microplasma is a locally very restricted process, and it will always be extremely sensitive to any modification. Zener references are very stable if a high number of microplasmas determines the reference voltage, what can be obtained with a high zener current. The LT1236 has the lowest possible current consumption for a Zener. Therefore, only a limited number of microplasmas determines the reference voltage. I made experiments with static magnetic fields in the chip plane of the LT1236, and switching of different microplasmas can be oberved when the orientation of the magnetic field is rotated. For instance, the RTN is largely modified with rotating the field (besides a weak Hall effect). Unfortunately, these effects are not stable. The instability cannot be reliably improved by magnetic fields. LT1236 advantages were: low consumption, good short term noise, hermetic package; disadvantage: temperature drift demands temperature control. As consequence, the LT1236 project was interrupted, and JFET references are tested now. Preliminary result: good candidate, but the compensated JFET reference will always be a very expensive solution, since its adjustment needs many hours of laboratory work.