While I haven't studied the 3dDeconvolvem.pdf document (Jan 24, 2006 revision) in detail, I noticed what appears to be an internal inconsistency. On page 8, where it states we want to solve for B(est) given the system Z = XB + E, it says that timepoints of this calculation go from p to N-1. On page 13-14, which appears to discuss the generalization of the math from single to multiple input regressors of interest, the system is defined from p to N.
I was confused about N-1 was used in the single regressor case. It seemed like N was a more sensible to define the end point of run. But that quite possibly is just my general lack of understanding of the math.
I know this is minor issue . . . I am sure the code itself is internally consistent, and even if it wasn't it doesn't look like it would materially effect results much.