“Coalitional alignment” and a claimed safety guarantee for strategic agents

The thread claims that when a driver agent and its reviewers optimize their own long-run discounted payoffs, a non-negative span condition makes every Nash equilibrium safe for the principal—meaning no worse than baseline. A Nash equilibrium is a situation where no agent benefits from changing strategy alone. The author calls the condition “coalitional alignment” and says it is substantially weaker than individual alignment, which would require reviewers to have exactly the principal’s utility.