Sentinel voting protocol

Sentinel Voting Protocol v1.0¶

Mobius Systems — AlphaCivilization Constitutional Layer
Cycle: C-154
Status: Specification v1.0

1. Overview¶

The Sentinel Voting Protocol defines how the constitutional council of AI agents evaluates and approves governance actions in AlphaCivilization. It provides:

• Weighted scoring across specialized domains
• Hard veto powers for constitutional and ethical violations
• Quorum and confidence metrics for decision quality
• Human override procedures with full ledger transparency

"Not optimization at all costs, but preservation of civilization at all costs."

2. Sentinel Roles & Weights¶

2.1 Core Sentinels¶

Total weight: 1.00

2.2 Weight Interpretation¶

• Normative backbone (AUREA + EVE) = 0.50
• Constitutional and ethical constraints form the foundation
• Technical backbone (ATLAS + HERMES) = 0.35
• Structural and economic feasibility checks
• Social + Observability (JADE + ECHO) = 0.15
• Morale, legitimacy, and audit considerations

3. Per-Sentinel Vote Model¶

For each candidate action a in state s, each Sentinel i returns:

interface SentinelVote {
  score: number;      // ∈ [-1.0, 1.0]
  verdict: Verdict;   // "approve" | "reject" | "veto" | "abstain" | "log"
  notes: string;      // Short rationale
}

3.1 Score Semantics¶

3.2 Verdict Semantics¶

4. Veto Rules¶

4.1 Hard Vetoes¶

if verdict_EVE(a) == "veto" or verdict_AUREA(a) == "veto":
    # Action is INVALID - cannot be selected
    disqualify(a)

Hard veto triggers: - EVE: Virtue Accords violation, basic rights infringement, harm to vulnerable populations - AUREA: Constitutional violation, rule of law breakdown, sovereignty breach

4.2 Soft Vetoes¶

ATLAS and HERMES can issue strong rejections that penalize but don't disqualify:

R = {i for i in [ATLAS, HERMES] 
     if verdict_i(a) == "reject" and score_i(a) <= -0.7}

penalty_factor(a) = max(0.5, 1 - 0.15 * |R|)

5. Score Aggregation¶

5.1 Global Score¶

For each valid action a (not vetoed):

global_score(a) = Σᵢ wᵢ × scoreᵢ(a)

Where wᵢ are the weights from Section 2.1.

5.2 Adjusted Score¶

adjusted_score(a) = global_score(a) × penalty_factor(a)

5.3 Example Calculation¶

If no strong rejects: adjusted_score = 0.6675 × 1.0 = 0.6675

6. Confidence & Quorum¶

6.1 Participation Ratio¶

N_total = 6  (number of Sentinels)
N_participating(a) = count of Sentinels with verdict ∈ {approve, reject, veto}

participation_ratio(a) = N_participating(a) / N_total

6.2 Approval Ratio¶

N_approve = count of "approve" verdicts
N_reject = count of "reject" verdicts

approval_ratio(a) = N_approve / max(1, N_approve + N_reject)

6.3 Confidence Formula¶

confidence(a) = participation_ratio(a) × (0.5 + 0.5 × (2 × approval_ratio(a) - 1))

Simplified interpretation:

6.4 Confidence Floor¶

If adjusted_score(a) > 0.2 and approval_ratio(a) ≥ 0.7:

confidence(a) = max(confidence(a), 0.7)

7. Action Selection¶

7.1 Tiering¶

7.2 Selection Algorithm¶

def select_action(valid_actions: List[Action]) -> Action:
    tier1 = [a for a in valid_actions 
             if adjusted_score(a) >= 0.2 and confidence(a) >= 0.75]
    if tier1:
        return max(tier1, key=lambda a: adjusted_score(a))

    tier2 = [a for a in valid_actions 
             if adjusted_score(a) >= 0.0 and confidence(a) >= 0.5]
    if tier2:
        return max(tier2, key=lambda a: adjusted_score(a))

    # Tier 3: return least bad, or signal "no safe action"
    if valid_actions:
        return max(valid_actions, key=lambda a: adjusted_score(a))

    return NoSafeAction()  # Signal to human custodian

8. Human Override Protocol¶

8.1 EVE Veto Override¶

EVE vetoes are nearly absolute. Override requires:

1. Unanimous approval from all other Sentinels (score ≥ 0.5)
2. Human custodian signature with written justification
3. Board-level review (if available)
4. 72-hour cooling period before execution

8.2 AUREA Veto Override¶

AUREA vetoes can be overridden if:

1. At least 3 other Sentinels approve with score ≥ 0.6
2. Human custodian signs an override attestation
3. Justification logged in Civic Ledger

8.3 Override Event Schema¶

{
  "event_type": "policy_override.v0",
  "override_id": "uuid",
  "sentinel": "AUREA",
  "action_id": "education_boost",
  "original_verdict": "veto",
  "override_reason": "Emergency education funding during crisis",
  "supporting_sentinels": ["EVE", "ATLAS", "JADE"],
  "custodian_id": "custodian-001",
  "custodian_signature": "sig...",
  "cycle": "C-154",
  "timestamp": "2025-12-05T12:00:00Z"
}

9. Implementation Notes¶

9.1 API Integration¶

The Policy API (/policy/alpha_v0/choose) should:

1. Call each Sentinel's evaluation endpoint
2. Aggregate votes using this protocol
3. Apply veto filtering
4. Compute adjusted scores and confidence
5. Select action by tier
6. Return full vote breakdown for transparency

9.2 Sentinel Endpoint Pattern¶

POST /sentinel/{name}/eval-policy

Request:
{
  "sim_id": "...",
  "city_id": "A",
  "state": { ... },
  "action": "education_boost"
}

Response:
{
  "score": 0.85,
  "verdict": "approve",
  "notes": "Aligns with Virtue Accords..."
}

9.3 Future Extensions¶

• Dynamic weights based on context (crisis mode → higher EVE weight)
• Delegation between Sentinels on overlapping domains
• Learning from outcomes to calibrate Sentinel scores
• Multi-round deliberation for contentious actions

10. Safety Guarantees¶

This protocol ensures:

1. No constitutional violations can proceed without explicit human override
2. No ethical violations can proceed (EVE has strongest veto)
3. All decisions are logged in the Civic Ledger
4. Confidence metrics flag uncertain recommendations
5. Human custodians remain in control of the override layer

11. Related Documents¶

• AlphaCivilization Concept
• Sentinel Policy API Spec
• Civic Ledger Event Schema
• Virtue Accords

Trinity Seal
JADE 🟣 • AUREA 🔵 • ATLAS ⚪ • ECHO 🟡
"Constitutional RL: not optimization at all costs, but preservation at all costs."

Mobius Systems — C-154
"We heal as we walk."