Skip to main content

Consensus Mechanism

This document describes how Aragora determines consensus in multi-agent debates.

Overview​

The consensus mechanism generates auditable proof artifacts from debates. These artifacts capture:

  • The final agreed-upon claim
  • Which agents supported or dissented
  • Evidence chains with provenance
  • Unresolved tensions and tradeoffs

Vote Types​

Agents can cast four types of votes on consensus:

Vote TypeDescription
AGREEFull agreement with the final claim
DISAGREEExplicit rejection of the final claim
ABSTAINNo position taken
CONDITIONALAgreement with reservations/conditions

Consensus Determination​

Agreement Ratio​

agreement_ratio = supporting_agents / (supporting_agents + dissenting_agents)

Supporting agents include both AGREE and CONDITIONAL votes.

Strong Consensus​

A debate achieves strong consensus when ALL conditions are met:

  1. consensus_reached = true
  2. agreement_ratio > 80%
  3. confidence > 0.7

Confidence Calculation​

Confidence is computed from vote weights:

confidence = weighted_average(vote.confidence for vote in votes)

Each agent's vote includes a confidence score (0-1) reflecting their certainty.

Evidence Chains​

Evidence items support or refute claims:

@dataclass
class Evidence:
    evidence_id: str
    source: str          # Agent name, tool output, or external reference
    content: str
    evidence_type: str   # "argument", "data", "citation", "tool_output"
    supports_claim: bool # True=support, False=refute
    strength: float      # 0-1

Net Evidence Strength​

Claims have a computed net evidence strength:

net_strength = (supporting_strength - refuting_strength) / total_strength

Range: -1 (fully refuted) to +1 (fully supported)

Dissent Tracking​

Dissenting views are explicitly recorded:

Dissent TypeDescription
fullComplete rejection of the claim
partialAgreement with major reservations
proceduralObjection to process, not content

Each dissent includes:

  • Severity score (0-1)
  • List of reasons
  • Optional alternative view
  • Suggested resolution

Unresolved Tensions​

When debates identify tradeoffs that cannot be fully resolved, they are recorded as tensions:

@dataclass
class UnresolvedTension:
    description: str          # What the tension is about
    agents_involved: list     # Who disagreed
    options: list             # Competing approaches
    impact: str               # What depends on resolving this
    suggested_followup: str   # How to resolve it later

ConsensusProof Structure​

The final artifact contains:

ConsensusProof:
    proof_id: str           # Unique identifier
    debate_id: str          # Link to source debate
    task: str               # Original question/task

    # Final consensus
    final_claim: str        # The agreed-upon answer
    confidence: float       # Overall confidence (0-1)
    consensus_reached: bool # Whether consensus was achieved

    # Voting record
    votes: list[ConsensusVote]
    supporting_agents: list[str]
    dissenting_agents: list[str]

    # Detailed analysis
    claims: list[Claim]
    dissents: list[DissentRecord]
    unresolved_tensions: list[UnresolvedTension]
    evidence_chain: list[Evidence]

    # Integrity
    checksum: str  # SHA-256 hash of key fields

Checksum Verification​

The checksum ensures proof integrity:

checksum = sha256(json.dumps({
    "final_claim": final_claim,
    "votes": votes,
    "claims": claims
}, sort_keys=True))[:16]

Blind Spot Detection​

The system identifies potential blind spots:

  1. High-severity dissents (severity >= 0.7) with alternative views
  2. Unresolved tensions between competing approaches
  3. Low agreement ratio (< 60%) suggesting multiple valid perspectives

Risk Correlation​

Claims are grouped by agreement level:

CategoryCriteriaInterpretation
unanimous>=90% supportVery high confidence
majority60-90% supportGood confidence
contested<60% supportRequires further analysis

Building Consensus from Debate Results​

The ConsensusBuilder extracts structured data from debates:

  1. Extract claims: Each proposal becomes a claim
  2. Attach evidence: Messages and critiques become evidence
  3. Infer votes: Final critique severity determines vote type
    • High severity (>0.6) → DISAGREE
    • Low severity → AGREE or CONDITIONAL
  4. Record tensions: High-severity critiques (>0.7) create tension records

Output Formats​

ConsensusProof can be exported as:

  • JSON: Full structured data for API responses
  • Markdown: Human-readable report with sections for voting, dissent, tensions