Frontier Council

Multi-model deliberation for important decisions. 4 frontier LLMs debate a question, then Claude judges and synthesizes.

Inspired by Andrej Karpathy's LLM Council, with added blind phase (anti-anchoring), explicit engagement requirements, rotating challenger role, and social calibration mode.

Models

Council (deliberators):

• GPT (gpt-5.2-pro)
• Gemini (gemini-3-pro-preview)
• Grok (grok-4)
• Kimi (kimi-k2.5)

Judge: Claude Opus 4.5 (synthesizes + adds own perspective)

Installation

pip install frontier-council

Or with uv:

uv tool install frontier-council

Setup

Set your OpenRouter API key:

export OPENROUTER_API_KEY=sk-or-v1-...

Optional fallback keys (for flaky models):

export GOOGLE_API_KEY=AIza...      # Gemini fallback
export MOONSHOT_API_KEY=sk-...     # Kimi fallback

Usage

# Basic question
frontier-council "Should we use microservices or monolith?"

# With social calibration (for interview/networking questions)
frontier-council "What questions should I ask in the interview?" --social

# With persona context
frontier-council "Should I take the job?" --persona "builder who hates process work"

# Multiple rounds
frontier-council "Architecture decision" --rounds 3

# Save transcript
frontier-council "Career question" --output transcript.md

# Share via GitHub Gist
frontier-council "Important decision" --share

# List past sessions
frontier-council --sessions

All sessions are auto-saved to ~/.frontier-council/sessions/ for later review.

Options

How It Works

Blind First-Pass (Anti-Anchoring):

1. All models generate short "claim sketches" independently and in parallel
2. This prevents the "first speaker lottery" where whoever speaks first anchors the debate
3. Each model commits to an initial position before seeing any other responses

Deliberation Protocol:

1. All models see everyone's blind claims, then deliberate
2. Each model MUST explicitly AGREE, DISAGREE, or BUILD ON previous speakers by name
3. After each round, the system checks for consensus (3/4 non-challengers agreeing triggers early exit)
4. Judge synthesizes the full deliberation

Rotating Challenger:

• One model each round is assigned the "challenger" role
• The challenger MUST argue the contrarian position and identify weaknesses in emerging consensus
• Role rotates each round (GPT R1 → Gemini R2 → Grok R3 → Kimi R4...) to ensure sustained disagreement
• Challenger is excluded from consensus detection (forced disagreement shouldn't block early exit)

Anonymous Deliberation:

• Models see each other as "Speaker 1", "Speaker 2", etc. during deliberation
• Prevents models from playing favorites based on vendor reputation
• Output transcript shows real model names for readability

When to Use

Use the council when:

• Making an important decision that benefits from diverse perspectives
• You want models to actually debate, not just answer in parallel
• You need a synthesized recommendation, not raw comparison
• Exploring trade-offs where different viewpoints matter

Skip the council when:

• You're just thinking out loud (exploratory discussions)
• The answer depends on personal preference more than objective trade-offs
• Speed matters (council takes 60-90 seconds)

Python API

from frontier_council import run_council, COUNCIL
import os

api_key = os.environ["OPENROUTER_API_KEY"]

transcript, failed_models = run_council(
    question="Should we use microservices or monolith?",
    council_config=COUNCIL,
    api_key=api_key,
    rounds=2,
    verbose=True,
    social_mode=False,
)

print(transcript)

License

MIT