Guess how well the typical top scorers in the class did on this one.
A well-written document would be a great illustration of Haskell-style lazy evaluation, though
The one I got was very explicit: Read the entire test first, and only after you've read all the questions and instructions are you to begin work.
Let's just say that I failed and I was pissed about it, without any particularly good basis or justification, other than being 12-ish years old.
Yes, that is the assumption. Many mathematical models make assumptions like that. It seems much more likely than the opposite assumption which is that they are independent.
Since you're attempting to model human behavior you are going to have to make assumptions that are ultimately incorrect. That's how humans process information. We join the dots and infer things in countless ways that mean we could be incorrect, but which none-the-less are better than wildly guessing (http://en.wikipedia.org/wiki/List_of_cognitive_biases). It would be probably better, if you are trying to model this, to think of it as a machine learning classifier. Brown M&Ms existing would be a prominent feature, not proof.
Second, let's imagine that M&M:s and safety is handled by the same people. Then brown M&M:s could easily be negatively correlated with equipment errors because a competent technician would spend more time double-checking equipment than full-filling bullshit demands.