The ideal system for machines, for humans, and for the agents that are now both is: strict at the boundary congruent in what it sends capability-scoped in what it trusts fast to fail loudly instead of corrupting silently and brave enough to run exactly one deliberate vulnerability loop with the node it has chosen