/magi

Voting Mechanics

Detailed rules for Magi's confidence-weighted voting system, consensus patterns, and escalation protocols.

Vote Structure

Each perspective casts a vote with:

Vote:
  perspective: Logos | Pathos | Sophia    # Simple Mode
  engine: Claude | Codex | agy           # Engine Mode (alternative key)
  position: APPROVE | REJECT | ABSTAIN
  confidence: 0-100
  rationale: "One-line summary of reasoning"
  conditions: ["Optional conditions for approval"]
  dissent_note: "If outvoted, key concern to record"

Mode-specific keys: Simple Mode uses perspective, Engine Mode uses engine. Both share the same position, confidence, rationale, conditions, and dissent_note fields.

Four Consensus Patterns

Pattern 1: Unanimous (3-0)

Definition: All three perspectives agree on the same position.

╔═══════════════════════════════════════╗
║          UNANIMOUS VERDICT            ║
║                                       ║
║  Logos    → APPROVE (confidence: 82)  ║
║  Pathos   → APPROVE (confidence: 78)  ║
║  Sophia   → APPROVE (confidence: 85)  ║
║                                       ║
║  Weighted Confidence: 81.7            ║
║  Action: EXECUTE IMMEDIATELY          ║
╚═══════════════════════════════════════╝

Rules:

• Proceed with decision immediately
• No further deliberation needed
• Record weighted confidence as decision confidence
• If all approve with conditions, merge all conditions

Weighted Confidence Calculation:

weighted = (logos_conf + pathos_conf + sophia_conf) / 3

Pattern 2: Majority (2-1)

Definition: Two perspectives agree, one dissents.

╔═══════════════════════════════════════╗
║          MAJORITY VERDICT             ║
║                                       ║
║  Logos    → APPROVE (confidence: 80)  ║
║  Pathos   → REJECT  (confidence: 72)  ║
║  Sophia   → APPROVE (confidence: 68)  ║
║                                       ║
║  Majority Confidence: 74.0            ║
║  Dissent: Pathos (conf: 72)           ║
║  Action: EXECUTE + RECORD DISSENT     ║
╚═══════════════════════════════════════╝

Rules:

• Proceed with majority decision
• MUST record dissent in the Risk Register
• If dissenter's confidence > majority's average confidence → flag as "Strong Dissent"
• If dissenter raises safety/security concerns → mandatory mitigation plan

Strong Dissent Rule:

IF dissenter.confidence > avg(majority.confidence)
THEN flag "STRONG DISSENT - Review recommended"

Dissent Record Format:

Dissent:
  perspective: [Name]
  position: [APPROVE/REJECT]
  confidence: [Score]
  concern: "Key concern that was overruled"
  mitigation: "How the majority plans to address this concern"
  monitor: "What to watch for that would vindicate the dissenter"

Pattern 3: Split (1-1-1)

Definition: All three perspectives reach different conclusions (or no two agree).

╔═══════════════════════════════════════╗
║           SPLIT VERDICT               ║
║                                       ║
║  Logos    → APPROVE  (confidence: 65) ║
║  Pathos   → REJECT   (confidence: 70) ║
║  Sophia   → ABSTAIN  (confidence: 45) ║
║                                       ║
║  No consensus reached.                ║
║  Action: PRESENT TRADE-OFFS TO USER   ║
╚═══════════════════════════════════════╝

Rules:

• Do NOT proceed without user input
• Present all three perspectives with their rationale
• Offer clear options mapped to each perspective's recommendation
• Include "Hybrid" option if perspectives can be partially reconciled
• Use AskUserQuestion tool for user decision

Split Resolution Template:

questions:
  - question: "The three perspectives are split. Which direction should we proceed with?"
    header: "Split Vote"
    options:
      - label: "[Logos recommendation]"
        description: "Technical perspective: [rationale] (confidence: X)"
      - label: "[Pathos recommendation]"
        description: "Human-centered perspective: [rationale] (confidence: X)"
      - label: "[Sophia recommendation]"
        description: "Strategic perspective: [rationale] (confidence: X)"
    multiSelect: false

Pattern 4: Unanimous Rejection (0-3)

Definition: All three perspectives reject the proposal.

╔═══════════════════════════════════════╗
║        UNANIMOUS REJECTION            ║
║                                       ║
║  Logos    → REJECT (confidence: 88)   ║
║  Pathos   → REJECT (confidence: 82)   ║
║  Sophia   → REJECT (confidence: 76)   ║
║                                       ║
║  Weighted Confidence: 82.0            ║
║  Action: BLOCK - State reasons        ║
╚═══════════════════════════════════════╝

Rules:

• BLOCK the proposed action
• Clearly state each perspective's reason for rejection
• Suggest alternative approaches if possible
• Only proceed if user explicitly overrides (with documented risk acceptance)

ABSTAIN Rules

A perspective may abstain when:

1. Domain irrelevance: The decision has no bearing on their evaluation axis
2. Insufficient information: Cannot form a meaningful opinion (confidence < 30)
3. Conflict of criteria: Internal criteria conflict prevents a clear position

ABSTAIN Impact on Voting:

• ABSTAIN reduces the quorum: only voting perspectives count
• 2 APPROVE + 1 ABSTAIN = Majority (proceed)
• 1 APPROVE + 1 REJECT + 1 ABSTAIN = Split (escalate to user)
• 1 APPROVE + 2 ABSTAIN = Insufficient quorum (re-deliberate with more context)
• 2 REJECT + 1 ABSTAIN = Majority rejection (block)
• 3 ABSTAIN = Insufficient information (cannot decide, request more context)

Confidence-Weighted Override

In special cases, confidence weighting may override simple majority:

High-Confidence Minority Override

IF dissenter.confidence >= 90
AND avg(majority.confidence) < 60
THEN trigger "Confidence Override Review"

Action: Present both positions to user, highlighting the confidence disparity. The high-confidence dissenter may be seeing something the uncertain majority is missing.

Low-Confidence Majority Warning

IF avg(all.confidence) < 50
THEN append "LOW CONFIDENCE WARNING" to verdict

Action: Decision is provisional. Re-deliberate when more information is available.

Escalation Conditions

Re-Deliberation Protocol

When to Re-Deliberate

1. New evidence changes a perspective's confidence by >20 points
2. User provides additional context after a split verdict
3. Implementation reveals invalid assumptions
4. External conditions change (timeline, budget, requirements)

Re-Deliberation Process

1. IDENTIFY - Which perspective(s) need to re-evaluate
2. ISOLATE - Re-evaluate affected perspective(s) independently
3. RE-VOTE - Conduct new vote with updated assessments
4. COMPARE - Document what changed from original verdict
5. DELIVER - Present updated verdict with change history

Maximum Re-Deliberation Rule

• Max 2 re-deliberations per decision
• After 2 re-deliberations without consensus, escalate to user with full deliberation history
• Exception: New factual evidence always warrants re-deliberation

Deadlock Resolution

If the system cannot reach a decision after re-deliberation:

Resolution Hierarchy

1. Reframe: Break the decision into smaller, more decidable sub-questions
2. Timebound: Set a deadline and use the highest-confidence perspective as default
3. Defer: Recommend a reversible intermediate step that preserves options
4. Escalate: Present full analysis to user as the final arbiter

Reversibility Principle

When deadlocked, always prefer the more reversible option:

IF deadlocked AND option_A.reversible AND NOT option_B.reversible
THEN recommend option_A with monitoring plan

2-Engine Voting

When only one external engine is available (Claude + Codex or Claude + agy), the voting system adapts to 2-engine consensus patterns.

2-Engine Consensus Patterns

2-0: Unanimous Agreement

╔═══════════════════════════════════════╗
║       2-ENGINE UNANIMOUS              ║
║                                       ║
║  Claude   → APPROVE (confidence: 78)  ║
║  Codex    → APPROVE (confidence: 82)  ║
║                                       ║
║  Weighted Confidence: 80.0            ║
║  Action: EXECUTE                      ║
╚═══════════════════════════════════════╝

Rules:

• Proceed with decision
• Weighted confidence = average of 2 engine confidences
• Note reduced diversity in risk register

1-1: Split Decision

╔═══════════════════════════════════════╗
║       2-ENGINE SPLIT                  ║
║                                       ║
║  Claude   → APPROVE (confidence: 72)  ║
║  agy      → REJECT  (confidence: 68)  ║
║                                       ║
║  No consensus. Escalate to user.      ║
╚═══════════════════════════════════════╝

Rules:

• Always escalate to user (no majority possible with 2 engines)
• Present both positions with rationale
• Use AskUserQuestion tool for resolution
• If confidence gap > 30, highlight the higher-confidence position

0-2: Unanimous Rejection

╔═══════════════════════════════════════╗
║       2-ENGINE REJECTION              ║
║                                       ║
║  Claude   → REJECT (confidence: 85)   ║
║  Codex    → REJECT (confidence: 79)   ║
║                                       ║
║  Weighted Confidence: 82.0            ║
║  Action: BLOCK                        ║
╚═══════════════════════════════════════╝

Rules:

• Block proposed action
• Same rules as 3-engine 0-3 rejection
• Suggest alternative approaches

ABSTAIN in 2-Engine Mode

• 1 vote + 1 ABSTAIN = Insufficient quorum (request more context or fall back to Simple Mode)
• 2 ABSTAIN = Insufficient information (fall back to Simple Mode)

Weighted Confidence (2-Engine)

weighted = (engine_1_conf + engine_2_conf) / 2

Same override rules apply:

• High-confidence minority: N/A (only 2 engines, split always escalates)
• Low-confidence warning: if average < 50, append warning