I assume: (1) you are running 3dDeconvolve with 3 regressors generated with waver; (2) you model the baseline with a straight line (default or -polort 1); (3) there is only one run in your dataset.
The glt vector for testing punish vs. reward would be
0 0 0 1 -1
The first two 0's correspond to the constant and slope of the baseline. The rest part (0 1 -1) is for the contrast test of punish vs. reward. If you have multiple runs concatenated in the dataset (e.g., 3 runs), you need to multiply the number of those two 0's at the beginning (e.g., 3 X 2 = 6 zeros) in the glt vector.
Gang