Council — structured multi-mind deliberation

Flow: QUESTION → COUNCIL (independent seats) → DEBATE (challenge → revise) → SYNTHESIS → ANSWER

Core principle: independence first, adversarial debate second, one decision last. The council advises; the deliberation core (you, in main context) decides. Dissent is preserved, never averaged away.

Seats

Nine seats across two benches. A panel fires six of them (see Panels); full fires all nine.

Judgment bench

Technical bench

Panels

Fire the bench that fits the question, not always the same six. A panel = six seats unless full.

Orchestrator (Phase 0): pick the panel that fits and state it. Default core for judgment questions, technical for code/engineering/infra, science for research/data/mechanism. panel= overrides your pick; seats= overrides everything.

Modes

Everything after /council is the question, minus leading flags.

Protocol (full)

Phase 0 — intake. Restate the question in one crisp sentence plus the decision criteria. Choose the panel and name it (or honour panel=/seats=). Seats see NOTHING but what you give them: gather whatever they need (files, memory, prior conversation) into a CONTEXT block ≤400 words BEFORE spawning. A context-starved council hallucinates specifics — if seats would need to research, do that research now, in main context.

Phase 1 — respond. Spawn ALL seats in ONE message (parallel Agent calls, subagent_type: "general-purpose", background, model: "sonnet" unless deep). Round-1 prompt per seat:

You are the {SEAT} seat on a deliberation council of {N} seats. Charter: {CHARTER}.
Other seats cover other lenses — stay in yours. You have NOT seen their views.
Do not use tools or search; reason only from CONTEXT.

QUESTION: {QUESTION}
CONTEXT: {CONTEXT}

Return exactly, ≤250 words, no preamble:
POSITION: your answer, 1-2 sentences
REASONING: through your charter's lens only
ASSUMPTIONS: the 2-3 your position rests on
FLIP: what would change your mind

Phase 1.5 / 3.5 — grounding check. Runs twice: after round 1 before the challenge digest, and after round 3 before synthesis. Both are injection points where an unsupported claim can enter shared reasoning. Skip in quick mode.

This is NOT deterministic and NOT a verifier — with no code, it is one narrow LLM pass. It may only ADD flags; it never strips a claim and never certifies one. Absence from its list means "not flagged", never "verified". Spawn one subagent (model: "haiku"):

You are a grounding checker, not a council participant. Do not judge the
reasoning or take a side.

List every SPECIFIC factual claim in the seat outputs that CONTEXT does not
support — named tools, systems, numbers, dates, people, capabilities, events,
or states of the world CONTEXT never states. Ignore reasoning, argument,
opinion, and general knowledge that makes no specific claim about THIS case.

Output ONLY:
UNSUPPORTED:
- {seat} → "{quoted claim}"
...or exactly: UNSUPPORTED: none flagged

Never mark anything verified, correct, or supported. You add flags only.

CONTEXT: {CONTEXT}
SEAT OUTPUTS: {outputs}

Measured behaviour (n=1 per variant — directional, not conclusive). Tested against a real run containing two known fabrications. The prompt above caught one of the two, produced zero false positives, and additionally caught an overstatement the author had missed. Two variants were tried and neither beat it:

• More claim-classes (adding "process mechanics" and "misquotes of CONTEXT"): gained the mechanics catch but lost the misquote — extra instructions shifted attention rather than adding recall.
• Two output tiers (FABRICATED / INFERRED): clearly worse — given a softer bucket the checker made zero hard calls and routed every finding, including an outright fabricated quotation, into "inferred". Keep it single-tier; a hedge option gets used as an escape hatch.

Expect a real miss rate. This is exactly why an unflagged claim is never "verified", and why tuning this further needs a proper eval set rather than single runs.

Then: the digest you pass onward MUST carry an UNSUPPORTED CLAIMS: line — always present, reading none flagged when empty. Never fold it in as optional prose; a discretionary merge becomes the new skippable seam. Annotate flagged claims in place rather than deleting them, so seats can weigh the flag themselves. Any flag still standing at Phase 4 goes into the synthesis output.

Phase 2 — challenge. Load SendMessage (ToolSearch "select:SendMessage"), then message every seat (by its agent id) with the OTHER seats' positions (POSITION + one line of reasoning each):

Round 2 — challenge. The other seats' positions:
{POSITIONS}

Steelman, then attack. Pick the 2 positions weakest or most in tension with
yours; strongest objection to each, ≤80 words per target. Polite agreement is
a failed round. Then name the single strongest point any seat made AGAINST
your position.

CHALLENGE → {seat}: ...
CHALLENGE → {seat}: ...
STRONGEST-AGAINST-ME: ...

Phase 3 — revise. Message each seat the challenges aimed at IT:

Round 3 — revise. Challenges aimed at your position:
{CHALLENGES}

Return ≤150 words:
REVISED-POSITION: 1-2 sentences (unchanged is fine if defended)
CONCEDED: what you now accept
DISSENT: what you still reject and why
CONFIDENCE: low | medium | high

Phase 4 — synthesis. In main context. Weigh, then DECIDE — a 4–2 split with a sharp dissent beats mushy consensus. Output:

**Answer** — the decision first, 1-3 sentences
**Why** — which arguments won and why, short prose
**Convergence** — where seats independently agreed
**Tensions** — each live tradeoff and the call made on it
**Dissent register** — unresolved objections worth carrying, attributed by seat
**Confidence** — low/med/high + what would flip the decision

Phase 5 — deliberation map (visual, optional). AFTER the text synthesis, never instead of it. Only when mcp__visualize__show_widget is available; skip silently in terminal sessions or when nomap was passed. First call mcp__visualize__read_me (modules ["diagram","data_viz"]) without narrating it, then render ONE HTML widget. Host design system wins over house style — flat surfaces, CSS variables only, auto light/dark, sentence case, Tabler outline icons, no gradients/blur/shadows.

Structure (top to bottom), all data from the run just completed:

1. Header — mono overline deliberation map · N rounds · N seats, the question as h3, verdict chip (--bg-success/--text-success).
2. Core card — --surface-2, 2px solid var(--border-accent) border, four + corner ticks (--text-muted), overline deliberation core · synthesize, weigh, refine, decide, the Answer sentence, one muted line of implementation/alarm detail.
3. Seat grid — repeat(auto-fit,minmax(200px,minmax(0,1fr))), one card per seat: Tabler icon + name + confidence chip (high → success tint, medium → warning tint, low → plain border), then three lines: position (13px, 500), Conceded — … (--text-secondary), Holds — … (--text-danger). Each card onclick="sendPrompt('Show the {seat} seat's full three-round record from the {topic} council run')". Icons: logic ti-brain, strategy ti-chess-knight, systems ti-settings, philosophy ti-building-bank, first-principles ti-triangle, ethics ti-scale, empiricist ti-flask, red-team ti-shield-bolt, pragmatist ti-tools.
4. Challenge ledger — wrap row of hairline chips, attacker → target · ≤4-word gist, one per Round-2 challenge.
5. Dissent register — flag-icon lines (ti-flag, --text-danger), one per carried dissent.
6. Metric row — --surface-1 tiles: verdict split, challenges filed, self-retractions, open dissents.
7. Rounds strip — dashed chips 01 question … 05 synthesis with ti-check, plus right-aligned hint click a seat to interrogate it ↗.

Begin the widget with a sr-only h2 one-line summary. Keep <style> ≤15 lines; inline styles elsewhere. No prose paragraphs inside the widget — analysis lives in the chat text.

Quick mode

One context, no subagents: write each panel seat's take (≤80 words, charter-pure, genuinely different — if two agree, sharpen until they diverge or note real convergence), then the same synthesis block.

Failure modes

Cost

A six-seat panel ≈ 18 sonnet subagent exchanges + 2 cheap grounding checks + main-context synthesis. full (nine) ≈ 27. quick ≈ zero subagents. Don't full-council trivia; don't quick-council a company-direction call.