A few weeks we saw 18 bit DAC noise in DARM go away when switching from state 1 to state 3 (medium noise to low noise, alog 64212).
Nominally, we operate our PUM drivers in state 1.  

Our DACs were changed from 18 bit to 20 bit recently (alog 64274).

Today, I rechecked our DAC noise situation.  I revamped the code in /ligo/gitcommon/labutils/coil_drivers_state_switch which Anamaria helped me write last year.

The code starts with the PUMs in their low noise state.  In our case I switched them all to state 3 for this test.
Then the code flips each quad PUM coil driver to the high noise state 2, waits for 300 seconds, and switches back to the low noise state 3.
Finally, the code uses DARM to estimate the excess power in state 2, 
computes the analog filters engaged on the coil drivers for state 3,
and applies that TF to the excess power to make a projection of the coil driver noise.

Results
Unlike in 2021, this time we saw clear noise in DARM when the coils were in state 2.
This makes making the projections much easier, but PUM DAC noise may be closer to DARM than in 2021.

Plot 1 is the comparison of a reference DARM in state 3 vs the high noise coils in state 2.
Plot 2 is the comparison of a reference DARM in state 1 (nominal) vs the high noise coils in state 2.
Plot 3 is a direct comparison over a longer timescale of state 1 vs state 3 (15 minutes starting from 1344376944 and 1344378599, respectively).

Switching rapidly between Plot 1 and 2, it's easy to see that state 3 is much better.  
State 1 DAC noise is only a factor of 2 below DARM between 20 and 30 Hz.

The direct comparison plot between state 1 and 3 is a bit more sobering, similar to Evan's plot here.  
We can't see the improvements around the calibration lines because I forgot to turn them off.
The other improvements are more nebulous: maybe some improvement at 25 and 30 Hz, but nothing crazy.

I think if we can swing it (which we have been this lock), running in state 3 should be fine.  
This would put our DAC noise around 20x lower than DARM.

It seems that Livingston was unable to do this last year due to PUM saturation issues alog 54727.
Our MASTER_OUTs seem to be far lower than Livingstons were.
If necessary we can apply a dither line to our DACs as a test, but this projection predicts that would have no effect on DARM.