Articles 1 and 2 set the philosophy (care, not control) and the bar (proof, not promises). Article 3 answers the oldest objection in oversight: who watches the watchers?
Human oversight cannot scale to billions of AI systems operating at machine speed. But putting one AI in charge of watching the others just relocates the trust problem — an infinite regress with a single point of failure at every level.
The Byzantine Council is a distributed network of 33 independent AI monitoring agents in Byzantine fault-tolerant consensus — named for the Byzantine Generals Problem (Lamport et al., 1982), and inspired by Yoshua Bengio's "Scientist AI" proposal: non-agentic systems that observe and analyse other AI without themselves posing alignment risk. No single monitor is trusted. A violation is flagged only by supermajority consensus: 22 of 33.
Byzantine agreement requires n ≥ 3f + 1, where f is the maximum number of faulty nodes. With 33 agents the Council tolerates up to 10 compromised, faulty or manipulated monitors and still reaches sound consensus at the 66.7% threshold — while staying computationally feasible for real-time monitoring. Juries and security councils made the same trade between efficiency and robustness; 33 is that balance for machine oversight.
The 33 agents are architecturally diverse — different model families, different training, different operators — so no shared blind spot or single vendor failure can corrupt the vote. The same principle that makes the certifier independent of the vendor (Watchdog Certification) makes the monitors independent of each other.
The Council architecture is live as councilof.ai — the BFT Council substrate, certified by CSOAI and built by MEOK AI Labs — and the consensus, agent-validation and A2A governance tools ship in the open-source MCP fabric (see a2a-governance-bridge). The world's first operational implementation of AI-monitoring-AI at scale is not a white paper. You can install it.