Thanks for the quick response.

Actually, it's correctly guessing local times:

++ Auto-catenated datasets start at: 0 188 376 564 752 940 1128 1316
++ Input polort=3; Longest run=376.0 s; Recommended minimum polort=3 ++ OK ++
++ -stim_times using TR=2 s for stimulus timing conversion
++ -stim_times using TR=2 s for any -iresp output datasets
++ [you can alter the -iresp TR via the -TR_times option]
++ ** NOTE ** Will guess GLOBAL times if 1 time per line; LOCAL otherwise
++ ** GUESSED ** -stim_times 1 using LOCAL times
++ ** GUESSED ** -stim_times 2 using LOCAL times
++ ** GUESSED ** -stim_times 3 using LOCAL times

I wouldn't have thought that rounding would make a non-trivial difference, but it apparently does in my case. The reason I even tried manipulating this is because I was trying to figure out why -stim_file seemed to be doing a better job. It happens with multiple subjects. The images I posted are of the same 9 voxels in the same subject in the same condition, with the only difference being the version of the 1D file I used.

I'd be inclined to think it's just noisy beta values, but 1) it's always the first time point in the IRF that looks funny and 2) the IRFs look like hemodynamic response functions (in active regions) when I get rid of the decimals or use the -stim_file method.