Quorum

A 13-seat deliberative evaluation workflow within Guildhall. Expands de Bono’s Six Thinking Hats into 12 perspectives + 1 facilitator, designed for product ideation by someone who tends to assume the best in people.

Origin

The original six hats under-serve the space between “what could go wrong” (black hat pessimism) and “how will people actually behave when incentives shift” (systems thinking about emergent behavior). The Black Hat expansion into Gamer / Freeloader / Adversary captures three distinct failure modes that most product thinking collapses into a single bucket.

Structured as 12 perspectives + 1 facilitator, mapping to quorum math (halves, thirds, quarters, tenth-man tiebreaker).

The three phases

Phase 1: Fast sync kickoff (~1-2 minutes)

All 13 seats get the prompt simultaneously and produce their initial 2-3 sentence take. Small/fast models, parallel execution. The Facilitator collects these into a brief identifying early consensus, tension points, and seats that need deeper analysis.

Phase 2: Async deep dives (let it cook)

Only seats that flagged substantive concerns spin up heavier models for real analysis. Inter-department reports flow between seats — the Historian’s research feeds the Gamer’s exploit analysis; the Empath’s user-feeling analysis feeds the Evangelist’s recommendation framing.

Routing logic: if a seat’s Phase 1 output contains uncertainty markers (“potentially,” “depends on,” “worth investigating”) or direct flags to other seats, it qualifies for a deep dive. Otherwise its initial take stands.

Phase 3: Final decision meeting (sync)

The Facilitator reconvenes with all deep dive reports. The Tenth Man receives the emerging consensus and builds mandated dissent against it. Final synthesis produces the output report.

The 13 seats

Facts & Evidence (White Hat expansion)

Seat	Role	Model	Why
1 — Historian	What does data from analogous systems tell us?	Qwen3-30B-A3B	MoE efficiency, strong factual retrieval + RAG
2 — Surveyor	What data are we missing? What would most change our confidence?	Cogito 14B	Strong meta-reasoning about information gaps

Emotion & Experience (Red Hat expansion)

Seat	Role	Model	Why
3 — Gut	Immediate emotional reaction. Don’t justify it.	Qwen3.5-9B (GPU 1)	Fast, instinctive, different character than analytical seats
4 — Empath	How will users feel? Where does delight live? Frustration?	Qwen3.5-27B	User-perspective reasoning, vision-capable

Optimism & Value (Yellow Hat expansion)

Seat	Role	Model	Why
5 — Builder	Best case, what does this enable that didn’t exist before?	Qwen3-30B-A3B	Shares model with Historian (orthogonal perspectives)
6 — Evangelist	Why would someone champion this? What’s the recommendation sentence?	Mistral 7B Instruct (GPU 2)	Punchy persuasive framing, fast

Creativity & Alternatives (Green Hat expansion)

Seat	Role	Model	Why
7 — Inventor	What lateral ideas does this spark? What adjacent problem could it solve?	Qwen3.5-35B	MoE architecture for less predictable outputs
8 — Remixer	What does this combine, replace, or make obsolete?	Qwen3-30B-A3B	Structured reasoning, shares model with Historian/Builder (orthogonal)

Risk & Exploitation (Black Hat expansion)

Seat	Role	Model	Why
9 — Gamer	How will rational self-interested actors exploit this? The cobra effect seat.	Cogito 14B (deep thinking)	Shares model with Surveyor (orthogonal); hybrid reasoning with visible chain-of-thought
10 — Freeloader	Where does this create asymmetric value extraction? Does the system survive a 90/10 lurker ratio?	DeepSeek R1 14B	Unique model — systems/incentive reasoning. In tension with Evangelist, must be different model.
11 — Adversary	Who is actively incentivized to undermine, copy, or kill this?	Qwen3-8B	Competitive analysis, different model from Evangelist (catches different blind spots)

Process & Meta (Blue Hat expansion)

Seat	Role	Model	Why
12 — Tenth Man	Argue against whatever consensus the other 11 reached. Structurally mandated dissent.	DeepSeek R1 32B	Transparent reasoning chain. Must be different model from anything that produced the consensus.
13 — Facilitator	Synthesize. Where is conviction strongest? Uncertainty highest? What to validate next?	Qwen3-32B or Llama 70B	Strongest dense local model for synthesis across 12 outputs

Model reuse rules

Models can be reused across seats when:

• The seats ask fundamentally different questions (orthogonal perspectives)
• The seats are NOT meant to catch each other’s mistakes

Models must be unique when:

• Two seats are in direct tension (e.g., Evangelist vs Freeloader)
• A seat needs to challenge output that another seat produced (e.g., Tenth Man vs all)

Current reuse clusters:

• Qwen3-30B-A3B: Seats 1, 5, 8 (Historian/Builder/Remixer — orthogonal lenses)
• Cogito 14B: Seats 2, 9 (Surveyor/Gamer — orthogonal domains)

Key questions per seat

• Gamer: “If I assume everyone will do what’s easiest and most rewarding for them personally, does this still work?”
• Freeloader: “If 90% of users consume and 10% contribute, is this still viable?”
• Adversary: “Whose lunch money am I taking, and what will they do about it?”

Quick mode

If you only have 5 minutes, hit Seats 4 (Empath), 9 (Gamer), 10 (Freeloader), and 12 (Tenth Man). These surface things an optimistic builder misses.

Integration with Guildhall

Quorum is a workflow configuration within Guildhall. Each seat is a folder with a SEAT.md defining its role, model, system prompt, required context, and inter-seat dependencies. The Guild orchestrator reads these folders, spins up the right Ollama models, routes context, and collects outputs.

“I don’t know,” “it depends,” and “it is hard to predict” are valid seat outputs. Quorum values honest uncertainty over confident fabrication.

Relationship to Lugh

Quorum is not part of Lugh’s runtime pipeline. It is Lugh’s design review board, invoked at high-stakes decision points:

• Stage 0 curriculum review (before committing to a syllabus)
• Gate calibration (tuning pass/fail thresholds)
• Course Zero design (the critical thinking course that affects all subsequent courses)