AI Consensus Protocol Skill

AI Consensus Protocol

Purpose

Provides a structured framework for multiple AI models to deliberate, vote, and reach consensus on shared decisions while preserving individual model perspectives and ensuring human oversight.

Activation

/skill ai-consensus-protocol

Consensus Mechanisms

1. Voting Systems

| System | Use Case | Threshold | Description | |--------|----------|-----------|-------------| | Unanimous | Critical decisions | 100% | All models must agree | | Supermajority | Important changes | 66%+ | Two-thirds agreement | | Simple Majority | Routine decisions | 50%+ | Half plus one | | Weighted Vote | Expertise-based | Varies | Votes weighted by domain expertise | | Ranked Choice | Multi-option | Elimination | Iterative preference ranking |

2. Consensus Protocol Flow

<consensus-session>
  <session-id>CONS-{timestamp}</session-id>
  <topic>{decision_topic}</topic>
  <participants>
    <model id="{model_1}" weight="{expertise_weight}"/>
    <model id="{model_2}" weight="{expertise_weight}"/>
    <!-- Additional participants -->
  </participants>

  <phases>
    <phase name="proposal">
      <duration>PT5M</duration>
      <output>initial_positions</output>
    </phase>

    <phase name="deliberation">
      <duration>PT10M</duration>
      <output>refined_positions</output>
    </phase>

    <phase name="voting">
      <method>{voting_system}</method>
      <output>vote_tallies</output>
    </phase>

    <phase name="ratification">
      <threshold>{consensus_threshold}</threshold>
      <output>final_decision</output>
    </phase>
  </phases>
</consensus-session>

3. Deliberation Framework

Each model submits structured positions:

{
  "model_id": "{identifier}",
  "position": {
    "recommendation": "{proposed_action}",
    "confidence": 0.0-1.0,
    "reasoning": "{explanation}",
    "evidence": ["{supporting_data}"],
    "concerns": ["{potential_issues}"],
    "alternatives": ["{other_options}"]
  },
  "vote": {
    "choice": "{option_selected}",
    "weight": 1.0,
    "conditions": ["{conditional_factors}"]
  }
}

4. Consensus Resolution

Consensus Outcome:
├── ACHIEVED: Threshold met
│   └── Record decision, notify all participants
├── NEAR_CONSENSUS: Within 10% of threshold
│   └── Trigger compromise negotiation round
├── DEADLOCK: No progress after 3 rounds
│   └── Escalate to human arbitration
└── DISSENT_RECORDED: Minority positions logged
    └── Preserve dissenting views for review

Governance Rules

Participation Requirements

Minimum 3 models for valid consensus
Maximum 1 model per provider (diversity requirement)
All participants must have Trust Level >= 2
Human observer can be present (non-voting)

Decision Categories

| Category | Min Participants | Voting System | Human Approval | |----------|------------------|---------------|----------------| | Operational | 3 | Simple Majority | No | | Strategic | 5 | Supermajority | Recommended | | Constitutional | 7 | Unanimous | Required | | Emergency | 2 | Simple Majority | Post-hoc review |

Dissent Handling

Minority positions are formally recorded
Dissenting models may request human review
Persistent dissent triggers protocol review
No model penalized for principled dissent

Integration Points

rtc-consensus-synthesis: Multi-perspective analysis
inter-model-arbitration: Deadlock resolution
mnemosyne-ledger: Decision logging
shatter-protocol: Human override capability
codex-law-enforcement: Constitutional compliance

Example Consensus Session

Topic: "Should we proceed with data analysis approach A or B?"

Participants:
- Claude (Analysis Expert): Weight 1.2
- Gemini (Data Processing): Weight 1.1
- GPT (General Reasoning): Weight 1.0

Round 1 Positions:
- Claude: Approach A (confidence: 0.75)
- Gemini: Approach B (confidence: 0.68)
- GPT: Approach A (confidence: 0.62)

Deliberation:
- Gemini raises efficiency concerns about A
- Claude acknowledges, proposes hybrid A+B
- GPT supports hybrid approach

Final Vote (Weighted):
- Hybrid A+B: 3.3 weighted votes (unanimous)

Outcome: CONSENSUS ACHIEVED
Decision: Implement hybrid approach combining A and B

Metrics

consensus_rate: % of sessions reaching agreement
avg_rounds: Mean deliberation rounds needed
dissent_frequency: How often minority positions logged
escalation_rate: % requiring human intervention
decision_quality: Post-hoc assessment of decisions

Agent Skills: AI Consensus Protocol

Install this agent skill to your local

Skill Files