I checked the spectra of CMS MIXER DAQ channels and compared with RF45 signals to confirm if CARM CMS noise is limiting the sensing noise or not.
Figure 1-4 show the spectra and coherence of K1:LSC-REFL_PDA1_RF45_I_ERR_DQ and K1:LSC-CARM_SERVO_MIXER_DAQ_OUT_DQ with the gain of CMS input of 20dB, 0dB, -20dB, and -31dB, respectively.
Left top graph shows the CARM CMS input equivalent noise of each signal.
If we use -20dB or -31dB (nominal of current OBSERVATION state), the sensing noise seems to be limited by CMS noise while it can be avoided by using 0dB or more.
So, to avoid CMS noise, it would be better to use more than 0dB following the precaution written in circuit diagram of CMS circuit (fig5).

I also checked K1:LSC-CARM_SERVO_MIXER_DAQ_OUT_DQ when ITMX/ITMY are misaligned and the other suspensions are aligned (state1), ITMX/ETMX/ITMY/ETMY are misaligned and the other suspensions are aligned (state2), ITMX/ETMX/ITMY/ETMY are misaligned and PRM is aligned to maximize REFL PDA1 DC values (state3).
Since PRM alignment to maximize REFL PDA1 DC is far from the recorded values (fig6), it might be a clipping somewhere, so the alignment on REFL PDs should be checked at some point.
Figure 7 shows the result.

REFL0, REF3, and REF6 show the spectra of state1, 2, and 3, respectively
Since REF0 spectrum around 10 Hz is larger than the others (prbably due to the interference from ETM reflection at REFL), we need to misalign ETMs to evaluate sensing noise of CARM by using single bounce of PRM.
Also, since spectra in state3 shows the lower noise than those in state2, current good PRM alignment at OBSERVATION state would be not optimized in terms of CARM sensing noise.