OK, I think I know what happened. My stimuli are variable lengths, which aren't dealt with gracefully by afni_proc.py, so I'll probably have to do a stim_times_AM1/dmBLOCK scheme as outlined elsewhere on the boards (http://afni.nimh.nih.gov/afni/community/board/read.php?f=1&i=34770&t=34770&v=f).
I think I've said this before, but I'll say again: It would be really fantastic for afni_proc.py to handle amplitude-modulated regression. (In the meantime, I'll write Python scripts to edit the output of my afni_proc.py files... which still beats the old "hand-code preprocessing and regression" method. Don't think that I'm not grateful!)