I like "stabilization". I've also been exerting to incremental constructions as "scaffolding", from ethics. Both of these seem to describe parts of what's going on when a model "talks it out".