Correct, AnalysisGrabber is definitely not as well documented as it should be. Indeed you’d have to know the background of the onset detection technique used in this plugin to fully understand how it works.
Would you give some link, reference papers, anything that would explain how it works? Maybe you implemented a known algorithm? Having the name of it would be a good start.
So sorry for the late reply, I guess this is what we get for providing free tools.
In short: for each filter band there are two envelope followers, one fast, one slow. The slow one has an offset, setting it generally at a higher level than the fast one, unless there is a rapid change in the audio level. In that case, the fast follower momentarily crosses the level of the slow follower. Whenever that happens, a message is sent out for the Lo, Mid or Hi Trigger entries.
In the AnalysisGrabber interface, the number boxes next to the envelope follower graphs represent, from top to bottom:
Slow envelope follower interpolation time
Fast envelope follower interpolation time
Offset between two envelope followers
The graphs themselves serve only as a visual indicator of the relative values.
To calibrate the values, I personally tend to play a simple loop that is representative of the important triggers I expect, then for every band I click the monitor button that solos it (the Lo, Mid or Hi buttons) and first tweak its filter settings to properly isolate the part of the audio I want to get triggers for. Mostly, the default settings of the envelope followers give me pretty good results. Otherwise, tweaking the follower speeds and offset for a filter band allows you to determine how quickly it triggers, in other words, how sudden the changes need to be to detect an onset.