The Council
A 33-seat oversight design whose independence must be demonstrated before it can govern any safety decision.
Status: DESIGN — not a live system. The latest published three-leg point experiment measured n_eff=1 at rho=1. The historical DR-0007 number is unbound because its cited result artifact is absent. The claimed resilience under failed or adversarial voters remains rejected; see the Refutation Ledger. Everything below is the design visualization, not measured production behaviour.
What is designed multi-agent review?
A supermajority (23 of 33) of the designed council is required for any safety decision, so no single agent decides the outcome. Effective independence is measured, not assumed.
Designed quorum: 23/33
The designed threshold is 70% agreement (23 of 33 agents). This quorum is a design target — its measured status is published on the Refutation Ledger (DR-0007).
Multi-Provider Diversity
The target roster spans multiple providers. The current experiment used three model lineages across two providers and behaved like one independent leg.
Fail-Safe Operation (designed)
A 23-of-33 threshold is only arithmetic. It does not provide a failure-resilience guarantee unless voter independence, identity, transport, and signatures are real.
Simulated — not a live council
Design visualization. Measured status: DR-0007
33-seat design · latest n_eff 1.00 of 3 · 2 providers tested
The roster below is a target architecture, not provisioned infrastructure. The latest measured sample was fully correlated, so no capture-resistance claim is made.
GPT-4o
OpenAI
Claude 3.5
Anthropic
Gemini Pro
Llama 3
Meta
Mistral Large
Mistral
Command R+
Cohere
Qwen 2
Alibaba
Designed Seats
23-seat target threshold
Agent Distribution
Council Responsibilities
The proposed roster assigns seats to specialized review roles. These are design responsibilities, not provisioned agents or comprehensive live oversight.
Compliance Validators
Designed to check evidence against selected regulatory mappings
Risk Assessors
Designed to propose risk levels for accountable human review
Incident Analyzers
Designed to help investigate reported incidents and possible causes
Watchdog Monitors
Design target: scoped monitoring of registered systems; not live
Human Oversight
In the proposed workflow, automated tools support scoped monitoring while accountable human reviewers retain judgment for critical decisions. That Council workflow is not live.
Human + AI Collaboration
Design target for scoped monitoring — not a live 24/7 service
AI proposes risk levels, humans validate high-risk cases
Proposed enforcement actions would require authorised human approval
