|
| 1 | +# PAD-004: Automated Liability Adjudication & Dynamic Insurance Pricing for Autonomous Agents |
| 2 | + |
| 3 | +**Publication Date:** January 03, 2026 |
| 4 | +**Author:** Vouch Protocol Maintainers |
| 5 | +**Subject:** The "AI Black Box" — Cryptographic flight recorders for liability attribution |
| 6 | +**Status:** Public Prior Art |
| 7 | +**License:** Apache 2.0 |
| 8 | + |
| 9 | +## 1. Abstract |
| 10 | + |
| 11 | +This disclosure places into the public domain a system and method for automated liability attribution in autonomous AI agents. It describes the use of a cryptographically signed "Chain of Intent" (similar to a flight data recorder) to programmatically resolve disputes, process insurance claims, and adjust real-time risk premiums for non-deterministic AI systems. |
| 12 | + |
| 13 | +## 2. Problem Statement |
| 14 | + |
| 15 | +Traditional liability frameworks fail for AI agents due to the "Black Box Problem." When an agent causes damage (e.g., deletes a production database or executes a bad trade), it is often impossible to distinguish between: |
| 16 | + |
| 17 | +1. **Malice:** The agent was hijacked by an attacker. |
| 18 | +2. **Hallucination:** The agent's stochastic model failed despite valid inputs. |
| 19 | +3. **Misalignment:** The agent technically followed a vague user instruction but produced a harmful outcome. |
| 20 | + |
| 21 | +Without a tamper-proof audit trail of the *intent* (reasoning) at the moment of action, insurers cannot accurately price risk or adjudicate claims. |
| 22 | + |
| 23 | +### 2.1 The Liability Gap |
| 24 | + |
| 25 | +``` |
| 26 | +┌─────────────────────────────────────────────────────────┐ |
| 27 | +│ AI Agent Action │ |
| 28 | +│ ↓ │ |
| 29 | +│ Damage Occurs ($100,000) │ |
| 30 | +│ ↓ │ |
| 31 | +│ ┌─────────────────────────────────────────┐ │ |
| 32 | +│ │ WHO IS LIABLE? │ │ |
| 33 | +│ ├─────────────────────────────────────────┤ │ |
| 34 | +│ │ • User who gave instructions? │ │ |
| 35 | +│ │ • Model provider (OpenAI, Anthropic)? │ │ |
| 36 | +│ │ • Agent framework (LangChain)? │ │ |
| 37 | +│ │ • Infrastructure provider? │ │ |
| 38 | +│ │ • External attacker? │ │ |
| 39 | +│ └─────────────────────────────────────────┘ │ |
| 40 | +│ ↓ │ |
| 41 | +│ IMPOSSIBLE TO DETERMINE │ |
| 42 | +│ (No cryptographic evidence) │ |
| 43 | +└─────────────────────────────────────────────────────────┘ |
| 44 | +``` |
| 45 | + |
| 46 | +## 3. The Novel Solution: Cryptographic Adjudication |
| 47 | + |
| 48 | +We disclose a method where an "AI Liability Policy" is enforced via a protocol (like Vouch) that binds actions to signed intent logs. |
| 49 | + |
| 50 | +### 3.1 The Adjudication Logic |
| 51 | + |
| 52 | +The system defines a "Smart Adjudicator" that processes the cryptographic log: |
| 53 | + |
| 54 | +| Case | Condition | Attribution | Policy Triggered | |
| 55 | +|------|-----------|-------------|------------------| |
| 56 | +| **A: The Breach** | Signature invalid OR chain broken | External Attacker | Cybersecurity Policy | |
| 57 | +| **B: The Alignment Failure** | Signature valid, intent contradicts outcome | Model Failure | E&O (Errors & Omissions) | |
| 58 | +| **C: The Negligence** | Signature valid, user authorized risky action | User Operator | Claim Denied | |
| 59 | + |
| 60 | +### 3.2 Adjudication Flow |
| 61 | + |
| 62 | +```python |
| 63 | +def adjudicate_claim(vouch_token: str, damage_report: dict) -> ClaimResult: |
| 64 | + """ |
| 65 | + Automated liability adjudication using Vouch Protocol. |
| 66 | + """ |
| 67 | + # Step 1: Verify cryptographic chain |
| 68 | + valid, passport = Verifier.verify(vouch_token) |
| 69 | + |
| 70 | + if not valid: |
| 71 | + # Case A: Signature invalid = Attacker |
| 72 | + return ClaimResult( |
| 73 | + attribution="EXTERNAL_ATTACKER", |
| 74 | + policy="CYBERSECURITY", |
| 75 | + payout=damage_report.amount |
| 76 | + ) |
| 77 | + |
| 78 | + # Step 2: Check intent alignment |
| 79 | + intent = passport.payload |
| 80 | + outcome = damage_report.actual_outcome |
| 81 | + |
| 82 | + if not is_intent_aligned(intent, outcome): |
| 83 | + # Case B: Model hallucinated/misaligned |
| 84 | + return ClaimResult( |
| 85 | + attribution="MODEL_FAILURE", |
| 86 | + policy="ERRORS_AND_OMISSIONS", |
| 87 | + payout=damage_report.amount |
| 88 | + ) |
| 89 | + |
| 90 | + # Case C: User explicitly authorized the action |
| 91 | + return ClaimResult( |
| 92 | + attribution="USER_OPERATOR", |
| 93 | + policy="NONE", |
| 94 | + payout=0, |
| 95 | + reason="User authorized risky action in signed intent" |
| 96 | + ) |
| 97 | +``` |
| 98 | + |
| 99 | +### 3.3 Dynamic Risk Pricing (The "Pay-How-You-Drive" for AI) |
| 100 | + |
| 101 | +We further disclose a method for real-time premium adjustment based on protocol usage: |
| 102 | + |
| 103 | +| Risk Factor | Premium Adjustment | Rationale | |
| 104 | +|-------------|-------------------|-----------| |
| 105 | +| Uses Identity Sidecar (PAD-003) | -20% | Keys isolated from LLM | |
| 106 | +| Enforces Chain of Custody (PAD-002) | -15% | Full audit trail | |
| 107 | +| Implements policy guardrails | -10% | Pre-signing validation | |
| 108 | +| No Vouch integration | +50% | Black box risk | |
| 109 | + |
| 110 | +### 3.4 Insurance Provider as Root Verifier |
| 111 | + |
| 112 | +The insurance provider acts as a "Root Verifier," rejecting coverage for any transaction that does not carry a valid Vouch proof: |
| 113 | + |
| 114 | +``` |
| 115 | +┌─────────────────────────────────────────────────────────┐ |
| 116 | +│ Insurance Policy │ |
| 117 | +├─────────────────────────────────────────────────────────┤ |
| 118 | +│ COVERAGE REQUIREMENTS: │ |
| 119 | +│ │ |
| 120 | +│ ✓ All high-value actions (>$1000) MUST include │ |
| 121 | +│ valid Vouch-Token with signed intent │ |
| 122 | +│ │ |
| 123 | +│ ✓ Chain of Custody required for delegated actions │ |
| 124 | +│ │ |
| 125 | +│ ✗ Unverified actions are NOT covered │ |
| 126 | +└─────────────────────────────────────────────────────────┘ |
| 127 | +``` |
| 128 | + |
| 129 | +## 4. Implementation Architecture |
| 130 | + |
| 131 | +### 4.1 Components |
| 132 | + |
| 133 | +1. **Intent Logger** - Captures and signs all agent decisions |
| 134 | +2. **Policy Engine** - Validates intents before signing |
| 135 | +3. **Claim Processor** - Automated adjudication |
| 136 | +4. **Premium Calculator** - Real-time risk scoring |
| 137 | + |
| 138 | +### 4.2 Integration with Vouch Protocol |
| 139 | + |
| 140 | +```python |
| 141 | +class InsuredAgent: |
| 142 | + def __init__(self, signer: Signer, policy_id: str): |
| 143 | + self.signer = signer |
| 144 | + self.policy_id = policy_id |
| 145 | + self.intent_log = [] |
| 146 | + |
| 147 | + async def execute_action(self, action: dict) -> Result: |
| 148 | + # 1. Sign the intent BEFORE execution |
| 149 | + vouch_token = self.signer.sign( |
| 150 | + payload={ |
| 151 | + "action": action, |
| 152 | + "policy_id": self.policy_id, |
| 153 | + "risk_assessment": self.assess_risk(action) |
| 154 | + } |
| 155 | + ) |
| 156 | + |
| 157 | + # 2. Log for insurance purposes |
| 158 | + self.intent_log.append(vouch_token) |
| 159 | + |
| 160 | + # 3. Execute with proof attached |
| 161 | + result = await self.perform_action(action, vouch_token) |
| 162 | + |
| 163 | + return result |
| 164 | + |
| 165 | + async def file_claim(self, damage: dict) -> ClaimResult: |
| 166 | + # Submit intent log for adjudication |
| 167 | + return await insurance_api.adjudicate( |
| 168 | + intent_log=self.intent_log, |
| 169 | + damage_report=damage |
| 170 | + ) |
| 171 | +``` |
| 172 | + |
| 173 | +## 5. Prior Art Statement |
| 174 | + |
| 175 | +By publishing this disclosure, we establish prior art for any system that uses cryptographic intent logging to automate: |
| 176 | + |
| 177 | +1. **AI Professional Liability Insurance claims** - Automated E&O processing |
| 178 | +2. **Smart Contract-based insurance payouts** - On-chain adjudication |
| 179 | +3. **Legal dispute resolution** - Cryptographic evidence for AI liability |
| 180 | +4. **Real-time risk pricing** - Dynamic premiums based on protocol compliance |
| 181 | +5. **Regulatory compliance** - Audit trails for AI governance |
| 182 | + |
| 183 | +### 5.1 Related Work |
| 184 | + |
| 185 | +| Concept | Source | Differentiation | |
| 186 | +|---------|--------|-----------------| |
| 187 | +| Usage-based insurance | Auto insurance | Applies to AI agent behavior | |
| 188 | +| Smart contract claims | DeFi insurance | Uses intent chains, not just transactions | |
| 189 | +| E&O insurance | Professional liability | Automated adjudication via crypto proofs | |
| 190 | + |
| 191 | +### 5.2 Claims Established as Prior Art |
| 192 | + |
| 193 | +This disclosure precludes patents on: |
| 194 | + |
| 195 | +1. "Cryptographic liability attribution for AI agents" |
| 196 | +2. "Intent-based insurance adjudication for autonomous systems" |
| 197 | +3. "Dynamic premium pricing based on AI agent security practices" |
| 198 | +4. "Flight recorder patterns for AI agent liability" |
| 199 | + |
| 200 | +## 6. Reference Implementation |
| 201 | + |
| 202 | +- **Repository:** https://github.com/vouch-protocol/vouch |
| 203 | +- **Reputation System:** `vouch/reputation.py` |
| 204 | +- **Related Disclosures:** PAD-002 (Chain of Custody), PAD-003 (Identity Sidecar) |
| 205 | + |
| 206 | +--- |
| 207 | + |
| 208 | +*This document is published as prior art to prevent patent assertion on the described concepts while allowing free use by the community under the Apache 2.0 license.* |
0 commit comments