My answer: who knows? We've given it a deliberately vague goal statement (even more vague than the last one), we've given it lots of admittedly contradictory literature, and we've given it plenty of time to self-modify before giving it the goal of self-modifying to be Friendly.

Humans generally manage with those constraints. You seem to be doing something that is kind of the opposite of anthropomorphising -- treatiing an entity that is stipulated as having at least human intelligence as if were as literal and rigid as a non-AI computer.