I re-checked the ALS noise in the following configurations:

• PRM is misaligned.
• Michelson is not locked.
• TRX/TRY is maintained at ~1.

1. Arm lengths are controlled using POX/POY as a sensor, and the ETMs as actuators [orange traces in Attachment #1].
   • EX laser frequency is locked to the arm cavity length using the end PDH servo.
   • ALS beat note frequency fluctuations are read out using the calibrated DFD channels.
   • In this config, the DFD outputs are the out-of-loop sensor.
2. Arm lengths are controlled using the ALS beat frequencies as a sensors [blue traces in Attachment #1]
   • The control is no longer in the XARM/YARM basis, but in the CARM/DARM basis.
   • The CARM actuator is MC2, the DARM actuator is an admixture of the ETMs (equal magnitude of output matrix element, opposite sign).
   • The calibrated POX/POY photodiodes are used as the out-of-loop sensor in this config.

The RMS noise sensed by POX/POY is ~20pm, which is somewhat higher than the best I've seen (maybe the arms are moving more at the time of measurement or the AUX PDH loops need a bit of touching up). But the orange traces in the top row of Attachment #1 are already ~x2 better than the equivalent traces from when we were using the green beams to make the beats. So it's hard to explain the 0-300 fluctuations in the arm powers when the CARM offset is reduced to 0 - i.e. the ALS noise is becoming worse as I reduce the CARM offset (= have more circulating power compared to the conditions of this test). I assume the transmission is Lorentzian, in which case even if we have 5x the CARM linewidth worth of ALS noise, we should see the arm power fluctuate between ~10 and 300. 

* I notice that a big jump in the RMS sensed by POX/POY comes from the 24 Hz peak, which is presumably the Roll mode coupling to length - maybe a ResG can make the situation better. The high frequency noise can also be probably rolled off better.