My bet is on lutorm, the point is that the original questions are ambiguous and if you remove some of the ambiguity you increase the chances of someone getting it right.

My theory is that those who got it right (at least most of them) didn't get it because they "formed a consistent model". But simply because they made the same assumptions people before them did when creating the language. But of course, would need more data to be more sure of anything.