rc-verdict

Multi-agent collective decision-making via LLM expert panels.

rc-verdict assembles a panel of LLM "experts" — each wearing a distinct expertise hat with its own behavioral disposition — and drives them through a structured vote → debate → converge protocol mediated by a single overseer agent. Instead of trusting one model's first answer, you get a verdict with calibrated confidence, recorded dissent, and a fully auditable deliberation trace. It is a modular Python library (not a framework): plug in OpenAI, Anthropic, HuggingFace, Ollama — or the built-in MockBackend — and call it from any project.

📖 How the protocol works → docs/PROTOCOL.md

Installation

pip install rc-verdict                  # base package (includes MockBackend)
pip install "rc-verdict[openai]"        # or [anthropic], [huggingface], [ollama]
pip install "rc-verdict[dev]"           # development: testing, linting, coverage

Requires Python ≥ 3.11. Until the first PyPI release lands, install from source: pip install git+https://github.com/spidey99/rc-verdict

Quickstart

Run a full deliberation locally — no API keys needed:

import asyncio, json
from rc_verdict import Panel, VerdictConfig
from rc_verdict.backends.mock import MockBackend

vote = json.dumps({"position": "approve", "conviction": "high",
                   "evidence_quality": "high", "reasoning": "Analysis complete.",
                   "dissent_points": []})
config = VerdictConfig(default_backend="mock", min_panel_size=3, max_panel_size=3,
                       max_token_budget=50000, pool_experts_ratio=1.0)
panel = Panel(config, backend_factory=lambda provider="mock", **kw: MockBackend(responses=[vote] * 3))

result = asyncio.run(panel.deliberate("A caching PR with 92% test coverage.",
                                      "Should we merge this PR?"))
print(f"Decision: {result.decision}  Conviction: {result.confidence.conviction.value}")

With a real backend it is one call — experts are selected and generated for your question automatically:

from rc_verdict import deliberate

result = await deliberate("Should we open-source our ETL tool?")  # uses OPENAI_API_KEY
print(result.decision)
print(result.trace.to_markdown(verdict=result.verdict))  # full deliberation trace

See examples/ for more — custom experts, multi-backend routing, and more.

Core Architecture

┌─────────────────────────────────────────┐
│              CALLER / HOST APP          │
│  verdict = await panel.deliberate(input)│
└────────────────┬────────────────────────┘
                 │
        ┌────────▼────────┐
        │    OVERSEER     │  Mediates all rounds
        │  (single agent) │  Controls escalation
        └────────┬────────┘
                 │
     ┌───────────┼───────────┐
     │           │           │
 ┌───▼───┐  ┌───▼───┐  ┌───▼───┐
 │Expert │  │Expert │  │Expert │   3-7 panelists
 │  #1   │  │  #2   │  │  #3   │   Diverse hats
 │ "hat" │  │ "hat" │  │ "hat" │   + dispositions
 └───────┘  └───────┘  └───────┘

Protocol Flow

1. Expert Selection — Overseer analyzes input context and selects relevant experts from:

   • A pool of pre-canned experts (YAML-defined domain specialists)
   • Dynamically generated experts — model-derived hats + dispositions tailored to the input

2. Vote Round — Each expert independently casts a confidence-weighted vote

3. Unanimity Check — All agree → verdict returned immediately (cheap path)

4. Debate Round — If not unanimous:

   • Each expert's reasoning is shared with all others
   • Experts process peer reasoning and update positions
   • Overseer may inject clarifying questions

5. Re-Vote — Experts vote again with updated reasoning

6. Convergence — If still stuck after N rounds:

   • Elimination — Remove the expert with lowest confidence
   • Supermajority — Accept at ≥80% agreement threshold
   • Overseer Override — Synthesize a verdict from all reasoning
   • Deadlock — Return structured disagreement with all positions

Key Differentiators

Research Foundation

Design informed by 30+ papers (2023–2026) on multi-agent deliberation:

• Debate-or-Vote (NeurIPS 2025): Voting alone captures most MAD gains; debate is a martingale without bias correction
• Diversity of Thought (2024): Diverse model families outperform N copies of the same model
• MARS (2025): Meta-reviewer pattern achieves MAD accuracy at 50% token cost
• Demystifying MAD (Jan 2026): Confidence-modulated debate + diversity-aware initialization are the two interventions that actually work
• MachineSoM (EMNLP 2023): LLM agents exhibit human-like social dynamics; personality traits affect collaboration

More Resources

• examples/ -- Runnable scripts for common use cases
• docs/PROTOCOL.md -- Protocol state machine and convergence strategies
• docs/ -- Signal capture protocol and design documentation
• LICENSE -- MIT

Status

🚧 Alpha — Core protocol working. All backends (OpenAI, Anthropic, HuggingFace, Ollama, Mock) implemented. See KICKOFF_PROMPT.md for the full implementation spec and RELEASE_CHECKLIST.md for the path to PyPI.

License

MIT