I've been reviewing the designs and messages about the output stages and low-pass filters, and there's something that I haven't seen mentioned here.
The DAC in DDS chips like the AD985x is what's called a "current steering" DAC; it runs off a constant current source (controlled by Rset) and it steers the constant current between IOUT and IOUTB.
What that means is that any impedance asymmetry in the connections to IOUT and IOUTB turns into an RF current in the chip's power supply. To terminate one of these into a resister and the other into a passive LPF means the harmonics rejected by the filter come out on the power supply. Maybe they are adequately decoupled there, maybe not. But it reduces the quality of the harmonic rejection possible, no matter how good the LPF is. If both IOUT and IOUTB see the exact same impedance curve, there is minimal RF current to decouple at the power supply.
This error appears in the manufacturer's evaluation boards, both Type 1 and Type 2 modules (I think?), in AN-557, in the DDS-60, and has been propagated to most other implementations.
Terry's PCB design looks good, the use of a transformer is a suitable approach; though not perfect for a signal generator due to the low frequency performance of the magnetics it's probably fine for the PHSNA.
However, with a fast op-amp (e.g. AD8008), you can feed both currents into a summing amplifier before the LPF and get the identical impedance on both DDS outputs. The summing amplifier then drives the LPF, and can also provide enough gain to drive the output also, killing two birds with one stone - and it works down to DC.
BTW, I came across this issue when researching different applications of the AD9959 quad-DDS, which has a 500MHz core (good to 200MHz output). Any asymmetric currents cause cross-talk between the channels, and some of the boards don't handle this correctly. The official EVB uses transformers (4 at $10 each!). The cheap module I bought is fully asymmetric, but some use fast op-amps (cheaper than transformers). One module even provides two channels with transformers and two with op-amps. But it shows that designers using AD9959 are aware of this issue, which also affects AD9850/1 to a lesser degree (no on-chip cross-talk).
Clifford Heath, VK3CLF.