The test measures a particular kind, one that could probably be simulated (AI) with relative ease (I'm not saying it's easy, but that what is involved is a series of tests, trials, of possible transforms, and then a checking of transforms that work for simplicity. It's looking for an aha! pattern.

Here's a 2010 Master's Thesis that does pretty well on it. I remember someone came up with a better algorithm in the last year, but I'm not finding it quickly.