Multi-Agent Council

The 33-seat council — a design, not a live system

A proposed multi-agent review architecture with human oversight. Its thresholds are design parameters; live independence and resilience under independent failures are not established.

Designed supermajority

The 23-of-33 threshold is a design parameter. It does not prove independent voters, capture resistance or resilience under independent failures.

Transparent & Auditable

Design target: each vote would be Ed25519-signed and SHA-256 hash-chained so its record can be checked offline.

Target: rapid review

The architecture aims to make scoped reviews fast, but no live decision-time claim is made.

How the Council Is Intended to Work

1

Analyst Submits Report

A certified Watchdog Analyst reviews an AI system and submits a compliance report with their recommendation (Compliant, Conditionally Compliant, or Non-Compliant). The report includes detailed findings, evidence, and remediation steps.

2

Target: independently operated review legs

The proposed workflow sends a scoped claim and its evidence to independently operated model/provider legs and retains raw responses. That 33-seat workflow is not live today; the latest three-leg experiment measured rho=1 and n_eff=1. See the source record.

3

Council Vote

The proposed vote would require at least 23 of 33 seats to agree (⅔ + 1 supermajority). Effective independence must be measured, not assumed.

4

Decision & Escalation

Under the design, a report that reaches the threshold would be published; otherwise it would be escalated for human review. Proposed vote records would be Ed25519-signed and SHA-256 hash-chained for offline verification against the published signer.

Why test this design?

Aims to Avoid Single-Seat Decisions

The design requires a 23-of-33 supermajority so no single seat decides an outcome. That threshold does not establish that the seats are independent.

Detects Malicious Behavior

The design would flag persistent outlier votes for human investigation while retaining legitimate dissent. It has not demonstrated resistance to malicious or correlated legs.

Balances AI Speed with Human Wisdom

The intended workflow uses automated review for repeatable checks and sends significant disagreement to accountable human reviewers.

Creates Transparent Accountability

The target vote record is Ed25519-signed and SHA-256 hash-chained with its seat ID. Independent external time-anchoring is planned, not yet live; the proposed chain would prove record integrity, not the wall-clock time of signing.

Technical Specifications

Design Parameters

Total Agents
33
Required for Consensus
23 (⅔ + 1)
Supermajority margin
up to 10 dissent
Operational status
Design only

Vote Options

Agree

The analyst's report is accurate, evidence is sufficient, and the recommendation is appropriate for the identified risks.

Disagree

The report contains errors, missing evidence, or the recommendation does not match the severity of identified risks.

Abstain

Insufficient information to make a determination. Requires additional documentation or clarification from the analyst.

Explore the Council design

Review the proposed workflow, then inspect the measurements that currently constrain its claims.