I went for broke and built the prototype of the latest version with all 3 changes: the voltage regulator, the extra LC filter and the removal of the OP amp. It takes a day for the oscillator to stabilize, which means that it's medium tau stability is still rather poor, but the low tau stability seems to have broken through 1E-11. The ADEV at 2 seconds is an astonishing 7.9E-12 at the moment and crosses 1E-11 at around 8 seconds. At the moment, it hits 4E-11 at around 300 seconds, but the DAC is still slewing towards a stable center, so that should be regarded as preliminary still. The phase data is still being disturbed by excursions, but when the frequency is stable, it's staying fairly easily inside a ±5E-11 corridor - about 2-3 times the observed frequency stability of the previous model.

The next step is to keep running it until the DAC stops moving and then gather a couple hours of ADEV data to characterize it better. Then, I'll remove the precision voltage regulator and run it again (this time warming it up shouldn't take as long), then remove the extra LC stage of the power supply and add back the regulator. I've already tested the removal of the op amp - that change is going to stay for sure.

The ADEV all 3 ways will give us what the ideal configuration is. But if it turns out (as I suspect) that the precision regulator makes the difference, then I have to justify ignoring the recommendation of the Connor Winfield application note. Jim suggested an experiment consisting of tuning the control voltage and power supply voltage separately to see whether the output frequency responds to solely changes in the CV or whether it responds to the ratio of CV to supply voltage. The application note's concerns about variable temperature compensation curves for the supply rails implies that it's the ratio that controls the frequency. If I decide to keep the precision regulator, then that would have to mean that I'm satisfied with the ability of GPS discipline to moot any differences in behavior of that sort between the buck converter and Vref regulator.

Tonight I'll build up one of the DOT variants. It will require a couple tweaks of the firmware, and it'll be interesting to see how the phase discriminator works with a less stable oscillator.