The Problem We're Solving
When you ask a single AI a question, you get one perspective. That AI might be confident, but it could also be wrong, biased, or missing something obvious. It's like asking one person for advice and taking it as gospel.
What the Council Does
The Council of Clifs asks six different AI "personas" the same question. Each persona has a different job:
- The Architect — thinks about structure, systems, and how pieces fit together
- The Devil's Advocate — actively looks for holes and challenges assumptions
- The Guardian — focuses on risk, safety, and what could go wrong
- The Innovator — explores creative and unconventional approaches
- The Strategist — considers long-term implications and competitive angles
- The Builder — thinks practically about execution and implementation
Why This Works Better
Imagine you're deciding whether to launch a new product. Instead of one opinion, you get six different lenses examining the same decision. When they agree, you have high confidence. When they disagree, you've found the interesting parts that need more thought.
The 10th Man Rule: When everyone agrees unanimously, an automatic "shadow auditor" kicks in to argue the opposite position. This prevents dangerous groupthink — the same principle Israeli intelligence uses for critical decisions.
The Result
You get a synthesized recommendation that accounts for multiple perspectives, highlights where the personas disagreed, and flags risks you might not have considered. It's like having a diverse advisory board that actually talks to each other.
Architecture Overview
The Council implements a multi-model consensus engine with parallel execution across heterogeneous LLM backends. Each persona maps to specific model configurations optimized for their cognitive style.
Model Routing
- The Architect → Claude Sonnet 4.5 (structured reasoning, low temperature)
- The Devil's Advocate → Claude Sonnet 4 (adversarial prompting)
- The Guardian → Claude Haiku (consistent risk heuristics)
- The Innovator → Llama 3.1 70B (diverse training distribution)
- The Strategist → Claude Sonnet 4 (game-theoretic reasoning)
- The Builder → Claude Haiku (practical implementation focus)
Consensus Mechanism
Responses are collected in parallel via AWS Bedrock. The system performs:
- Semantic similarity scoring — embeddings compared pairwise
- Divergence detection — flags when models disagree by >25%
- Weighted synthesis — persona weights configurable per domain
- Conflict arbitration — surfaces disagreements explicitly
10th Man Protocol: When initial response entropy falls below threshold (unanimous agreement), an async shadow process invokes an adversarial prompt against the consensus. This dissent analysis is injected before any escalation decision, implementing institutional checks against false confidence.
Infrastructure
Built on AWS Lambda (Node.js 20.x) with DynamoDB for session persistence. API Gateway handles routing via tw2c8v0s95. Async jobs support Opus 4.5 inference times (5+ min timeout). SHA-256 baseline tracking enables drift detection across runs.
Performance
Multi-model consensus achieves 90% task success rate vs 53-80% for single-model approaches. The architectural overhead (~3-5s parallel invocation) is offset by significantly reduced error rates and rework cycles.