That sounds pretty sophisticated. I like the idea a lot.

But for the sake of compute performance this would need some serious optimization or keep an extensive caching logic that keeps a users different trust scores and dynamically updates exactly those scores that are the oldest or have changed a lot.

This is just a suggestion from me, you could do it entirely different of course. Anyway, it's pretty fun to think about an architecture like this - will it slash fast if it looks like a users account was compromised? Will interaction with people with high trust scores ramp up the trust scores fast? So. Many. Things. To. Consider.