Likewise. I think for me there's quite a wide band of colours in the middle that I consider to be "neither/either", so I'm basically just picking a random answer for those.

A modified version of the test that finds two boundaries (green/neither/blue) could be interesting.

Or maybe it just needs to take more samples, in a more random order.

Same. Some of them are neither obviously blue nor obviously green, so what the test was measuring for me was what I was thinking about at the time, the decision I'd previously made, whether my mouse was currently hovering over "blue" or "green", etc.