I measured CSOFT P for an hour last night while at 10 W, using an excitation recording script in 
/ligo/gitcommon/noise_recorder/code/csoft_yaw_olg_injection_caller.py.
The measurement was taken with the RPC off (PNG 1), 
but the RPC is not doing very much anyway over the band at 10W (and also not at 60 W)  (PDFs 4 and 5)

The first three PDFs show the CSOFT P OLG with three binwidths: 0.1 Hz, 0.03 Hz, and 0.01 Hz.
The coherence thresholds are all different for each.
However, the 0.01 Hz one is the most interesting, as we see an obvious 2.2 Hz peak that was causing us problems locking this week.

I tried reducing the gain by a factor of 2, but this caused a 0.5 Hz ringup.  
It's hard to tell what madness is happening at 0.5 Hz, and likely the plant was changing slightly over this time, but there are clearly two peaks, and therefore some phase jumping.
Reducing the gain by a factor of two at 0.5 Hz probably put us in a place where this rapidly moving phase hurt us.

Instead, Elenna modified Evan's low pass filter to very slightly reduce the gain at 2.2 Hz and nowhere else (see PNG 1).
This is a better solution since it keeps the 0.5 Hz damping far above a gain of 1 at a location with rapid phase rotation.