They are the smallest and largest t-statistics found in each iteration of the randomization.

In turn, they are used to compute the 'XX.5percent.txt' output file, which is the threshold to use for voxelwise thresholding at the 5 percent (and other percentiles) global FPR level. This threshold is usually pretty strict, which is why clusterwise thresholding was invented in the first place. Due to spatial correlation (if blurring was used somewhere in the analysis), the Bonferroni correction by the number of voxels is usually WAY too strict. That is, taking the p-value voxelwise threshold to be 0.05/number-of-brain-voxels is overly conservative if spatial blurring were used. The XX.5percent.txt file gives the voxelwise t-threshold to use in that case -- this calculation is a side-benefit of the simulations needed for the cluster threshold calculations, and it is output at the request of the statistician-who-cannot-be-named.