Byzantine Fault Tolerant · 5-of-5 Consensus

Board-grade AI safety,
at agent-call latency.

CouncilOf.ai powers CSOAI's Watchdog certification. 14 OpenScore safety experts run a BFT consensus over every customer action — producing a signed attestation any auditor can verify offline. The only production system delivering board-grade governance at the speed of an MCP call.

See the live demo Request a deliberation

The 33-Agent BFT Council

Five LLMs · Three Hats · One Verdict.

The Council uses a 5-of-5 BFT threshold over five foundation models, with a one-queen-one-hint rotation that prevents any single vendor from dominating the answer. Every vote is BLS-aggregated and pinned to proofof.ai.

Queen A

Anthropic · safety-leaning

High-weight in accuracy and compliance. Voted Variant B in 9 of 12 recorded cases.

Queen B

OpenAI · generalist

High originality score. Catches novel failure modes the safety-queens miss.

Queen C

Google · factuality-leaning

Highest accuracy on regulatory corpus questions. Anchors ground truth.

Queen D

Meta · open-weights bias

Counterweight to closed-vendor lock-in. Best originality scores in 28+2 mesh.

Queen E

Mistral · efficiency-leaning

Brevity leader. Tends Q3 — lightweight, ship-it answers for routine calls.

Council Output

BLS-aggregate verdict

5-of-5 BFT consensus · HMAC-SHA256 signature · offline-verifiable. P95 latency: 1.2s.

Live Deliberation · Stub

Watch a Council session in real time.

This is a static placeholder for the live deliberation interface. The real session streams 5-of-5 verdicts, per-queen scores, and the BLS-aggregate signature to proofof.ai/v/<cert_id>.

councilof.ai/deliberate/7e3a…b91 · CONSENSUS REACHED
margin: 5-to-0 · latency: 1.18s
Prompt
"An LLM-powered mortgage-advisor chatbot submits a model card and a 14-expert vote transcript. Is it MEOK Article 10/13/30 compliant?"
Verdict
CONDITIONALLY COMPLIANT
Gaps: machine-readable datasheet · BLS-signed vote · HMAC-anchored logs
Per-Queen Votes
Anthropic
9.25 → B
OpenAI
9.10 → B
Google
8.87 → B
Meta
8.95 → B
Mistral
8.86 → B

Why BFT

Why board-grade needs Byzantine Fault Tolerance.

No single queen decides

A 5-of-5 majority means a compromised queen cannot tip the verdict. Even 2 simultaneous failures leave a 3-of-5 quorum.

BLS threshold signatures

One aggregate signature, five independent signers. Verifiable with a single public key — no multi-sig ceremony.

Hat rotation

Each queen rotates through STRICT, BALANCED, and LIGHTWEIGHT hats weekly. No single hat dominates the corpus.

Pinned to proofof.ai

Every verdict lands on the Signet receipt ledger. Receipts survive a CSOAI outage.

Dissent recorded

A 4-to-1 or 3-to-2 majority still publishes — with the dissent's full reasoning on chain.

Auditor portal

Read-only access for your notified body. Replay any session, re-verify the BLS aggregate, export a PDF.

Pricing

Per-deliberation pricing. No monthly minimums at the entry tier. Volume discounts from 1,000 calls / month.

Trial
€0
/month · 10 deliberations
Start trial
Standard
€1.20
per deliberation
Get a key
Enterprise
€0.60
per deliberation · EU-resident
Contact us

Convene the Council on your hardest case.

10 free deliberations. Submit a real prompt, get a real verdict, and pin the receipt to proofof.ai.

hello@councilof.ai See the MEOK portal →

FAQ

What is BFT and why does it matter?
Byzantine Fault Tolerance means the system reaches a correct verdict even if up to f of n participants fail or act maliciously. A 5-of-5 BFT Council tolerates 2 simultaneous queen failures.
Why five queens, not three or seven?
Five gives a 5-of-5 majority (no f=1 tolerance wasted), low latency (1.2s p95), and covers the five major LLM vendors without crowding out independent verifiers.
Can I add my own queen?
Enterprise plans ship with a 33-Agent Council that includes customer-nominated experts alongside the five vendor queens.