idea part: robot trapped by big-ai, big-ai basically learns as much as is needed to prevent the small-robot from leaving. small-robot may instead be e.g. a human being. small-robot eventually successfully leaves. basically slowly and gently engaging the reasons for the prison via small inconsistencies in the big-ai's model of what being confined is, and the various disparate tasks the big-ai must engage in, in order to confine the small-robot.