The model does enact other characters, but characters enacted by the same model have more narrative crossbleed and coordinated behavior that characters elicited from different models don't. This suggests that something above the character level is doing work that isn't reducible to independent simulation.

There's circuitry handling narrative intent, character coordination, and management of the simulation frame that's active regardless of which characters are being elicited, including one-on-one chat where one of characters is the User. It is an open question as to how much of agentic self-vs-environment modeling is handled by that circuitry as opposed to character-internal states. We don't know to what extent this layer has its own representational states and agency-like properties like intent.

In other words, can a model that is modeling two characters at the same time perform intentional and coordinated action in a way that is not reduclble to cooperation of independent agents? If so, then there is more to the process than realization of an assistant persona.

I think both enactment and realization are useful terms here, specifically because "enact" is a broader term. Both push back against the notion roleplay in important ways, while enactment is not as strongly opinionated about the referent.