Parliament：多智能体辩论共识引擎与合成话语系统

Parliament is an innovative multi-agent deliberation system inspired by human parliament debate mechanisms. It uses a five-role AI agent architecture (Proposer, Skeptic, Synthesizer, RedAgent, Sentry) to balance adversarial thinking and consensus达成. A key design feature is controlled AI misalignment via RedAgent to enhance reasoning resilience. The system supports local model backends like Ollama and LM Studio.

Parliament draws inspiration from human parliament debate. Unlike traditional multi-agent systems that pursue goal consistency, Parliament intentionally uses role对立—especially RedAgent's adversarial injection—to address group thinking risks and improve conclusion reliability. This design treats 'managed AI misalignment' as a core mechanism.

Five Roles:

• Proposer: Initiates debate with structured reasoning (default llama3.2).
• Skeptic: Challenges current positions (default mistral).
• Synthesizer: Integrates views to form consensus or mark irreconcilable splits (default qwen2.5, needs 0.7 confidence threshold).
• RedAgent: Adversarial injector every 3 rounds (default mistral-openorca) to test consensus robustness.
• Sentry: Monitors echo loops (via OSI) and convergence (default tinyllama).

Core Mechanisms:

• OSI (Opinion Shift Index): Quantifies view changes to detect echo chambers.
• Residue Score: Measures unresolved分歧 strength when splits occur.
• Model-aware scheduler: Groups same-model agents to minimize switching overhead.

System Architecture:

• TypeScript modular design with three core packages:

1. @parliament/core: DeliberationEngine, agent classes, model adapters (Ollama default, LM Studio, OMLX, OpenAI-compatible), OSI calibration.
2. @parliament/server: Hono-based REST API (POST /deliberate, GET /deliberate/:id, GET /health) with SQLite storage.
3. @parliament/cli: Local debate execution and history query.

Multi-Backend: Switch via PARLIAMENT_PROVIDER (ollama, lm_studio, omlx, openai_compatible) for offline/online use.

Parliament applies to:

1. Decision Support: Simulate stakeholder views to identify blind spots.
2. Creative Divergence: RedAgent breaks思维定势 for innovative ideas.
3. Argument Stress Testing: Pre-test policies/statements against objections.
4. Education/Research: Platform for studying consensus mechanisms and cognitive biases.

Limitations:

• Fixed 5 roles and max rounds (3 default).
• No dynamic agent addition/exit.
• OSI uses simple text similarity (misses semantic drift).
• No human interaction interface.

Future Plans:

• Add human supervisor role.
• Support dynamic agents and complex debate topologies.
• Improve OSI with advanced semantic models.
• Enable human-AI interaction.

Parliament redefines multi-agent systems by embedding structured adversarial collaboration instead of pursuing harmony. It borrows power separation ideas from democracy, using 'questioning' and 'opposition' as core components. For researchers and decision-makers, it offers a fully open-source, easy-to-deploy solution for high-quality deliberation and consensus.