I dont think you are understanding my experiment. The point isnt the topic. The point is that once you remove real world identifiers/context, the model drops safety hedging and becomes decisive.

Thats what happened with Alice/Bob (politics) and when I used fictional medical guidelines about a touchy subject. The mechanism is the same.

As far as I know, memories store tone and preference but wont override safety guardrails or political neutrality rules. Ill try it with a brand new account in a VPN later

"I would not expect that one party's candidate is always more correct over the other for every possible issue" --> I agree, just wanted to show the same test applied to a different side of the spectrum