(Edited this post; Forgot to account for the FMs other than 4 and 5... it now agrees better!)

I did some quick MATLAB simulation of the relevant loops to try and understand what was going on. I put the digital UGF around 200Hz, and then brought in the AO path with both signs. 

In these plots, blue is digital only, green is AO+digital with the crossover happening at the UGF, and red is the AO gain set to five times of what it was in the green curve. 

Based on the phase curves in the loop measurements, I would be inclined to say the pink -AO case corresponds to the opposite sign plot, and the +AO case to the same sign plot. 

This correspondence also holds for the appearance of the peaks in the noise curves, the Opposite sign case has a dip in loop gain at ~50Hz (pink curve, -AO), same sign around ~30Hz (brown curve, +AO). 

However, both of these look like they become unstable at some point in the transition! This agrees with our experience last night...

I'll fiddle around and try to come up with some compensating digital filter that will make the Opposite sign scenario work. 

The MATLAB code I used to make these plots is attached. 

clear all

cycleT = 60e-6;

% AI, AA shapes from http://nodus.ligo.caltech.edu:8080/40m/8555
[z,p,k] = ellip(4,4,60,2*pi*7570,'s');
AI = zpk(z,p,k*10^(4/20)) * zpk([],-2*pi*13e3,2*pi*13e3);
AI.OutputDelay = 1*cycleT;

[z,p,k] = ellip(8,0.001,80,2*pi*7570,'s');