The Board

Anti-sycophancy LLM council for business ideas.

Single-model LLM interactions are structurally biased toward agreement. RLHF training selects for responses that users rate positively — and users rate agreeable responses more positively. The result: any idea you bring to an LLM with enthusiasm gets received with enthusiasm. This is not a bug you can prompt your way out of. It requires a different architecture.

The Board runs five structured adversarial personas across multiple providers, has each model anonymously critique the others, and synthesizes the result — with critics by design, not by request.

Inspired by Karpathy's LLM Council and pAI's critical reviewer council (Poggio Lab, MIT).

---

How it works

Stage 1 — Parallel perspectives. Five personas run simultaneously, each on a configurable provider:

Persona	Default model	Role
The Advocate	openai/gpt-5.4-mini	Strongest honest case for
The Devil's Advocate	anthropic/claude-sonnet-4-6	Strongest honest case against + one fatal objection
The Realist	google/gemini-3.1-flash-lite-preview	Most likely actual outcome with base rates
The Market Skeptic	mistralai/mistral-large-3	Three unanswered market questions
The Execution Skeptic	deepseek/deepseek-chat	Three execution risks assuming market is correct

Stage 2 — Anonymous cross-review. Each persona reviews the others' outputs with identities stripped. Models evaluate reasoning, not source — preventing deference to "prestigious" models.

Stage 3 — Chair synthesis. A Chair model synthesizes into a fixed structure: where the board agrees, where it disagrees, the three hardest objections, what would change the picture, and concrete kill criteria. The Chair is not allowed to offer a verdict or end with encouragement.

---

Quickstart

# 1. Clone and install
git clone https://github.com/your-org/the-board
cd the-board
pip install -r requirements.txt

# 2. Configure
cp config.example.yaml config.yaml
# Edit config.yaml — add your OpenRouter API key
# Get a key at: https://openrouter.ai/keys

# 3. Run
python board.py

Or pass your idea directly:

python board.py --idea "A B2B SaaS tool for managing freelancer contracts"

---

Provider configuration

Each persona has a drop-in model field in config.yaml. Any OpenRouter model string works:

personas:
  devils_advocate:
    model: "anthropic/claude-opus-4-6"   # upgrade the most critical persona

  market_skeptic:
    model: "x-ai/grok-4"                 # swap to a different provider entirely

One API key. The Board uses OpenRouter to route to all providers. You need one key, not five.

Why provider diversity matters. When five different model families — trained on different data, with different RLHF lineages — produce the adversarial perspectives, the disagreements are real, not simulated. Running five GPT-5.4-mini calls produces one model wearing five hats. Running five different providers produces five genuinely distinct perspectives.

---

CLI options

python board.py [--idea TEXT] [--config PATH] [--fast] [--preview] [--budget USD]
                [--persona NAME:MODEL ...]

--idea TEXT           Idea to review (prompted interactively if omitted)
--config PATH         Path to config file (default: config.yaml)
--fast                Devil's Advocate + Chair only — lower cost, quick sanity check
--preview             Show cost estimate and model assignments, then exit
--budget USD          Auto-select cheapest models under this USD ceiling
--persona NAME:MODEL  Override one persona's model inline (repeatable)

Examples:

# Preview cost and model assignments before running
python board.py --idea "..." --preview

# Run with budget ceiling
python board.py --idea "..." --budget 0.15

# Override the Chair to use a stronger model
python board.py --idea "..." --persona chair:anthropic/claude-opus-4-6

# Quick sanity check (Devil's Advocate + Chair only)
python board.py --idea "..." --fast

---

Cost

Mode	Config	Approx. cost
Fast mode	Devil's Advocate + Chair	~$0.02–0.08
Full board, budget models	All → GPT-5.4 nano	~$0.02–0.05
Full board, recommended	Mixed providers (default config)	~$0.25–0.60
Full board, premium	All frontier models	~$1.00–2.50

Set cost_limit_usd in config.yaml to prevent accidental overruns. The board shows an estimate before running and aborts if it exceeds the limit.

---

Output

Each session saves a Markdown report to outputs/ with:

• The Chair synthesis (structured, fixed format)
• Full per-persona Stage 1 perspectives with model labels
• Full Stage 2 cross-review responses

outputs/
  20260411-143022-a-b2b-saas-tool-for-managing-fr.md

---

What The Board doesn't do

It does not eliminate sycophancy. Each individual persona is still an LLM with training-induced tendencies to soften criticism. The system prompts resist this but don't fully override it.

It does not replace judgment. The Board surfaces disagreement and names hard objections. A committed believer can still read the Kill Criteria and dismiss them. The tool changes the structure of the conversation — it cannot change the reader.

It does not give verdicts. The Chair is explicitly prohibited from saying "this is a good idea" or "this will fail." What it produces is: where confidence is warranted, where it isn't, and what would change that.

---

Customizing personas

Persona prompts live in personas/. Edit them or add new ones. When contributing a new persona, include:

1. The system prompt
2. Three example ideas it was tested on
3. The objections it produced (unedited)
4. Whether those objections surprised the proposer

A persona is better if it produces criticism that surprises — not just confirms what was already suspected.

---

Roadmap

• v0.1 — CLI, 5 personas, drop-in provider config, cost preview, Markdown output
• v0.2 — Web UI with provider dropdowns, domain templates (B2B SaaS, consumer, services), session history
• v0.3 — Track record mode, Slack/email integration, custom persona builder

---

The Board does not tell you whether your idea is good. It tells you what questions you haven't answered yet.