I'm not sure we will get any further. I'm sorry if things are unclear.

But I can't use the full time course.

Once we reach TR 27, live in the scanner, I want to instantly calculate the average (in an ROI) of TR 23,24,25,26.(index start at 0) and provde feedback to the subject. TR 23-26 is when the subject is watchig an image.

Best I can do is to use TR 0..26. Then, as you suggest, average within the ROI, and do 3dDeconvolve on this .1D file with 27 rows.
Then read TR avg(23, 24,25,26) as my image signal
and read TR (11,12,13,14) as my baseline signal (fixation cross).

Calculate the signal (image-baseline) / baseline.

Feed this value back to the fMRI computer and show the participant their activity while watching the image. Feedback will come at about TR 28/29. It takes about 2 seconds to process.

I tested this and it is not enough TR's. The increase in visual cortex is not completley gone due to the motion regressors, it's better than before (only using 13 TRs). But as you said I would probably need more TR's (about 50).. And since the first feedback is at 28 seconds (28TRs) I don't have more data to include in the regression. And for later feedbacks I don't want to base the regression on different ammount of data, to bias it.

Hope that makes it clear.

To me I only have two options. Include more time before first feedback in order for my motion regressors to not screw up my signal.
Or do motion correction offline afterwards and see if the real time feedback effects (learning how to down regulate) holds up.

About right? =)