After regenerating the NonSENS c-code and restarting the h1bos and h1oaf models yesterday, I tried some new cleaning during our commissioning time this afternoon. Now I can get all 3 of LSC, Jitter, and Laser noise subtraction all to work, with no mysterious minus signs needed.
Attached is my DTT 'NonSENS budget' from a time when we were still in commissioning. The blue is the CALIB_STRAIN_NOLINES channel, and red is CALIB_STRAIN_CLEAN. The black is the sum of all the other colors' noise estimates. This LSC subtraction was from a time before the MICH FF was retuned, so it may not have as big an effect were I to turn it on now that MICH FF was updated.
Today I turned on the Jitter and Laser noise subtraction during commissioning (without having retrained it), and it performed at a similar level to this plot from last week. The LSC subtraction was not as effective as this plot, but that's because the LSC in-loop feedforward was tuned later in the week last week, and so there just was less that needed to be subtracted.