Each engine alone is a part. The Council is where they work as a whole: one question runs through Memory → Permission → Context + Graph per agent, four members reason from their own scoped slices, and they deliberate — while you stay the sole decider. Many voices, one decider, every voice honest.
Personal memory is locked. Samples remain available.
Recalls related ground so you don't repeat a past move.
Proposes the forward move with the most momentum.
Names the risk — over-commitment or thin ground.
Weighs the evidence already in the vault.
Each agent gets a genuinely different scoped package — the Archivist sees more than the Scholar because their permission scopes differ. They argue from different evidence, so the disagreement has a basis. The same-model trap, answered structurally.
Every position is canCommit: false. The decision is null through the whole deliberation. No agent has a decide() path at all. Not "agents are asked to defer" — agents have no mechanism to act. Law 8.
Each view reports asserted vs derived connection counts. Honesty about what's stated versus what was inferred survives from the vault, through the graph, into the deliberation. Law 12.
In real model mode, each agent's position runs through the live Inference Proxy (/inference, OpenAI) — bounded by the application’s controls. In mock mode, a deterministic stance-flavored reasoner stands in so the deliberation is visible with no key. Because each agent carries its own scoped package, you get real diversity even on a single model — and full independence if you place different models behind different stances. The orchestration, scoping, and sovereignty are real now; only the text generator is optional.
The Council is the Kernel thinking out loud — many voices, one decider, every voice honest about what it knows and how it knows it.