Model Council

One prompt. Four frontier models. A structured debate. One synthesized verdict.

Model Council is a personal research tool that convenes four AI models, has them debate each other across 2–3 rounds, and uses Claude Opus 4.6 as a synthesizer to produce a final verdict with a side-by-side comparison table. It is built on top of OpenRouter so every model is reachable through a single API key.

The goal is simple: surface disagreement between frontier models on hard questions, rather than getting one confident answer from a single model and calling it a day.

---

The Council

Seat	Model	Role
1	anthropic/claude-sonnet-4.6	Debater
2	openai/gpt-5	Debater
3	google/gemini-3.1-pro-preview	Debater
4	x-ai/grok-4.20	Debater
Judge	anthropic/claude-opus-4.6	Synthesizer

All five are routed through OpenRouter. Swap a slug in app/lib/models.ts to change the lineup.

How a Debate Works

1. Round 01 — Independent. Every council model sees only the prompt and writes its own answer. Parallel calls.
2. Round 02 — Critique & Update. Every model now sees the other three Round 1 answers and is asked to identify where it agrees, where it disagrees, and update its own answer if the others have made valid points.
3. Round 03 — Final Statements (conditional). Claude Opus 4.6 classifies whether substantive disagreement remains after Round 2. If yes, a third round auto-fires — each model writes a final position. A "Force Round 3" checkbox lets you trigger it manually.
4. Synthesis. Opus 4.6 reads every round and produces a JSON verdict: a row per model (what they agreed with, what they held their ground on, their confidence, the arc of their argument), plus a synthesized final answer.

The UI streams the whole thing live over Server-Sent Events so you can watch tokens arrive in all four cards as the debate unfolds.

Web Search

Every model call can be augmented with OpenRouter's built-in :online mode, which uses Exa to inject fresh web results as context. A "⟡ Web Search" checkbox is on by default. Toggle it off for prompts that do not need current information — each searched call adds roughly $0.004 on top of the model's own token cost.

Synthesis and the disagreement check are never web-searched; they reason purely over the debate transcript.

The Verdict Table

Produced by the synthesizer in structured JSON (with a retry on parse failure). Rendered in the UI as:

Model	Agree	Disagree	Confidence	Reasoning Trace

Plus a final prose answer in hero type, and a one-click Export Markdown button that downloads a self-contained .md file with the prompt, every round's outputs, the verdict table, and the final answer.

---

Stack

• Framework: Next.js 16 (App Router) + React 19 + TypeScript
• Styling: Tailwind CSS v4 + Inter
• LLM client: openai SDK pointed at https://openrouter.ai/api/v1
• Streaming: native Web Streams + Server-Sent Events
• Animation: CSS keyframes + cubic-bezier(0.16, 1, 0.3, 1)

No database. No auth. Personal tool, runs on localhost.

Design

The UI follows a Bold Editorial Studio aesthetic:

• Strict black-and-white palette. Accent color only from model attribution.
• Inter at extreme weight and size contrasts. 700 headlines, 400 body, 14px uppercase monospace metadata.
• Custom 32-pixel difference-blend cursor that scales 2.5× over interactive elements; the native cursor is hidden.
• Asymmetric model cards with non-uniform border radii (100-pixel corner accents).
• Every hover transition runs at 700 ms or slower on cubic-bezier(0.16, 1, 0.3, 1). No system easing anywhere.

Full spec lives in PRD-model-council.md §7.4.

---

Quick Start

# 1. Clone
git clone <your-fork-url> model-council
cd model-council/app

# 2. Install
npm install

# 3. Add your OpenRouter key
cp .env.example .env.local
# Edit .env.local and paste your key from https://openrouter.ai/keys

# 4. Run
npm run dev
# Open http://localhost:3000

You need an OpenRouter account with credit on it. A typical debate runs roughly $0.10 to $0.40 in model costs depending on prompt size and how much each model writes; web search adds a few cents on top.

Project Layout

model_council/
├── PRD-model-council.md          # Full product spec (8 sections + design)
├── ASSUMPTIONS-model-council.md  # Risk scan across 8 categories
├── README.md                     # This file
└── app/                          # Next.js application
    ├── app/
    │   ├── layout.tsx            # Root layout: Inter, cursor, header, footer
    │   ├── page.tsx              # Hero, prompt bar, rounds, verdict, export
    │   ├── globals.css           # Bold Editorial base styles
    │   └── api/council/route.ts  # SSE streaming endpoint
    ├── components/
    │   ├── Cursor.tsx            # 32px difference-blend cursor
    │   ├── NavHeader.tsx
    │   ├── Footer.tsx
    │   ├── ModelCard.tsx         # Asymmetric radii per variant
    │   ├── RoundSection.tsx      # 2×2 grid per round
    │   └── VerdictTable.tsx
    └── lib/
        ├── models.ts             # Council + synthesizer config
        ├── council.ts            # Orchestration, prompts, JSON retry
        └── export.ts             # Markdown builder

Configuration

Env Var	Required	Description
OPENROUTER_API_KEY	Yes	Get one at https://openrouter.ai/keys
OPENROUTER_APP_NAME	No	Shown in your OpenRouter dashboard
OPENROUTER_APP_URL	No	HTTP-Referer header sent to OpenRouter

Swapping Models

Edit app/lib/models.ts:

export const COUNCIL_MODELS: CouncilModel[] = [
  {
    id: "sonnet",
    slug: "anthropic/claude-sonnet-4.6",
    displayName: "Claude Sonnet 4.6",
    provider: "Anthropic",
    shortLabel: "sonnet",
  },
  // ...
];

Any slug OpenRouter exposes will work. The synthesizer is defined just below and can be swapped the same way.

Known Limitations

• Self-reported confidence is not calibrated. When a model says "5/5", that does not mean it is actually right 95% of the time. Treat the numbers as relative, not absolute.
• Synthesizer bias. Using Opus 4.6 as judge when Claude Sonnet 4.6 is a debater is a real bias vector. Rotating the synthesizer occasionally is a good sanity check.
• Consensus is not truth. Four models can be wrong in the same direction. The verdict table shows what the council agreed on, not what is actually correct.
• No follow-up threading. One prompt, one debate, one verdict. Multi-turn refinement is not implemented in v1.

Roadmap

• Configurable council per run (swap models from the UI)
• Multi-turn follow-ups on the same debate
• Cost tracking in the footer
• localStorage-backed session history
• Custom system prompts per model

License

Personal project. MIT, effectively — do whatever you want, no warranty.

---

Built with Claude Code using the pm-skills plugin for the PRD and risk analysis.