Ask one model to decide a retool window "taking everything into account" and you get a fluent answer. It will mention output, safety, cost and the night shift, and it will settle them in a paragraph. What it will not do is argue. The functions were never held apart; they were blended in one pass, with whatever weights the model happened to carry that day, and the paragraph does not say which fact would have changed its mind.
Organizations learned long ago that this is how bad decisions are made. That is why a plant has a safety officer who does not report to production, why a company has a legal department that does not report to sales, why a crew has a flight director and a flight surgeon. The functions are kept apart on purpose, because an objection raised at full strength by someone whose only job is to raise it is the only way to find out what the proposal actually costs.
Mantikon keeps them apart in the same way. On the council the Critic can only attack, and only with a reproducible counter-argument; the Risk Guardian can only count what breaks; the Compliance role can only check the policy. None of them can pretend to be reasonable on another's behalf, and none of them can act. The proposal that survives all of them is not a compromise written by one hand; it is the thing that cleared every limit and still moved.
There is a second half to this, and it is the part the research on multi-agent debate keeps finding: roles that are merely prompted apart are not apart. Ten copies of one model share its training, its data and its blind spots, and a group of them will imitate each other into a confident, wrong consensus faster than one model alone. So the council measures its own independence: how correlated the roles' errors are, how many of the "confirmations" trace to one source, whether a confidence moved without new evidence. When independence is low, the record says so, and the decision is held or escalated rather than taken.
Two consequences follow. The first is that dissent is a result, not a failure: when a role cannot sign, the decision carries that fact, and the review date exists to find out who was right. The second is that the record of the argument is worth more than the verdict. A plant that can read why +12 became +7 with the window moved has learned something about itself. A plant that only receives "+7" has not.
That is why a decision needs a council, and why the council needs a protocol rather than a prompt.