Well, X and X'X are only composed of the elements of your stimulus files (plus some other parameters), therefore the problem would be something related to your stimulus files, but not the contrast setup.
Since you are really running 3dDeconvolve for linear regression, not deconvolution, the non-invertible problem is more likely to be some errors in one or more of those regressors. I suggest that you use "1dplot" to plot out all the 14 regressors (those *.1D files) and check whether you have mistakenly created two identical regressors. If you don't have a big screen, try the option "-use #" in 1dplot to plot only first # of data points in your regressors so that you could tell the differences among those curves.
As far as I can tell, you are doing fine with the contrast tests. One thing I am not so sure is that you might need to put those -glt lines before the output line "-tout -bucket gltreg" to make sure your output gltreg+orig.BRIK would include the differences and their corresponding T valuse of the constrast tests.
Let me know if this method helps you find out the problem.
Cheers,
Gang
