The way I read the code it tries a number of scans scripts by passing them to jpegtran one by one and remembering the size of the output.

The best result is then used.