I think the terms of the experiment are that Babylon will predict either Low or High.

I still think that if the predictive ability will come from deciding low or high, and two of six numbers don't count, then there's still just a 50% chance of being right.

if 3 & 4 are considered both high *and* low, then they're always right. We can discount them. However many 3s and 4s I roll will automatically be recorded as "correct", which will artificially inflate the predictive ability.

Sure, the "probability" that you'll either get a roll in the set of [1-4] or [3-6] is 2/3, but if we're testing the ability to predict the correct set, isn't there a 1 in 2 chance of predicting a roll that will actually be wrong?