Hi Misho,

It seems likely that something is not correct with the timing of the task, such as accounting for pre-steady state volumes.

Do the time series initially contain pre-SS volumes?
Do the stim times start at the beginning of pre-SS, or after?
Did you make any adjustments to the stimulus timing to account for pre-SS?
If you need to tell 3dTshift that the TR is 2s, maybe it is not correct in the headers, in which case 3dDeconvolve will not know either.
What does 3dinfo -tr show for those time series?

Smaller points:
I see you are telling afni_proc.py to remove the first 2 time points, which is only 4s of data. Typically it 7-10 s.
With TR=2s, such bandpassing will consume 60% of the degrees of freedom. It is generally not recommended, particularly in the case of task.
That might end up devouring some of the variance of those long, slow blocks, too.

- rick