The output leveling is looking pretty good now. According to an HP8484A/HP436A, the power level flatness of the TG alone is within about 0.3 dB p-p, from 10 MHz to 1500 MHz, with tilt compensation applied. It looks like the 8568's flatness (spec 2 dB p-p) is the limiting factor, so there's not much point to improving the TG any more. The adjustable tilt compensation works great, and can accommodate the 1 dB or so internal loss, and up to a couple meters of external DUT cabling. I've set up the variable power range for about >+2 to <-12 dBm, which in conjunction with the step attenuator, allows setting to any level from +2 to -82 dBm.
I have modified the original leveling system quite a bit, mostly to speed it up, get a wider power range, and incorporate the tilt compensation. The remaining stuff now is mostly about speed - as the variable power level is set toward the lower extreme, the detector diode signal gets very weak, and the response slows down, causing jagged amplitude on screen. When the 8568 sweep time is extended from the fastest 20 mSec, to around 50 or 100, everything smooths out nicely. I don't think the 8444A was intended to keep up with these kind of rates, so needs some fixing up.
Ed