What is Consilium
What is Consilium?
Consilium is a multi-AI agent deliberation platform. Unlike orchestration tools (CrewAI, AutoGen, LangGraph) that run models in parallel and pick one answer, Consilium implements formal debate protocols where AI models propose claims, challenge each other's reasoning with typed challenges, defend positions with categorized rebuttals, vote using social choice theory, and converge only when mathematically verified.
The result is a golden prompt (synthesized answer) with tracked confidence scores, dissent reports, and complete audit trails.
What Makes Consilium Different
1. True Deliberation, Not Orchestration
Orchestration runs models in parallel, picks the best. Deliberation makes them argue until the truth emerges. Models cross-examine each other with typed challenges (factual error, missing evidence, flawed logic) and respond with categorized rebuttals (concede, refute, qualify, redirect).
2. Formal Voting Theory
Not just "pick the most popular" — real social choice theory:
- Condorcet: finds the candidate that beats ALL others pairwise
- Borda count: confidence-weighted scoring
- Ranked Pairs: cycle-free tiebreaking
- Copeland: comparative win/loss analysis
3. Mathematical Convergence Detection
score = 0.4 × ranking_similarity + 0.35 × proposal_similarity + 0.25 × concession_rate
Debate continues until score ≥ 0.85 or max rounds reached. Not vibes-based — mathematically verified.
4. Mandatory Dissent Preservation
Agglomerative clustering identifies minority positions. Every result includes majority AND minority opinions. No decision is "unanimous" unless mathematically verified through clustering.
5. Confidence Calibration
calibrated = stability × (1 - concession_rate) × (1 - 0.3 × qualification_rate)
Models that cave under scrutiny get lower confidence scores. Based on "explanation stability" — do claims survive cross-examination?
6. Complete Audit Trail
Every phase recorded: model ID, input, output, tokens used, cost, latency. Full transparency into how consensus was reached. Required for regulated industries.
Platform in Numbers
Architecture
Web (Next.js 15, port 3000)
→ API (NestJS 11/Fastify, port 4000)
→ Agents (FastAPI/Python, port 8000)
→ LangGraph State Machine
├── PROPOSAL (independent analysis)
├── CHALLENGE (cross-examination)
├── REBUTTAL (defend/concede/qualify)
├── EVALUATION (rubric scoring)
├── VOTING (Condorcet → Ranked Pairs)
├── AGGREGATION (confidence-weighted)
├── CONVERGENCE (mathematical check)
└── OUTPUT (golden prompt synthesis)Database: PostgreSQL 16 (Neon) via Prisma ORM
Queue: Redis 7 (Upstash) + BullMQ
Auth: Clerk (JWT)
Streaming: Server-Sent Events (SSE)
Get Started
npm install -g @myconsilium/clinpm install @myconsilium/sdkpip install consilium