Council overview

The Council is a deliberate review mode. Local chat stays the fast default; the Council is for “review this decision / code / risk from several angles and show me the dissent.”

Three stages (Karpathy pattern)

User question
  │
  ├─▶ Stage 1 — First Opinions   parallel to N council seats
  │       (local: roles OR multi-model | online: 3–5 frontier models)
  │
  ├─▶ Stage 2 — Peer Review   each seat ranks the anonymized answers
  │       of the others by accuracy & insight (bias guard applies)
  │
  └─▶ Stage 3 — Chairman   synthesizes a reasoned final answer
          + Council trace (all opinions, ranking, dissent)

Each stage streams into the UI.

Why it works

• First opinions are produced independently, so seats don’t anchor on each other.
• Peer review surfaces where the seats disagree: the dissent is the signal, not noise.
• The Chairman must justify its synthesis against the visible opinions.

The Council trace shows every opinion, the peer ranking, and the marked dissent. Quality is enforced by a bias guard, and locally it runs as a local role council or local multi-model council.

Cost note (online): Stage 2 is roughly O(N²) in tokens. A per-query cost cap bounds it.