Agent Justice Protocol

Forensic investigation, dispute resolution, and risk assessment for the agent economy. The accountability layer (Layer 4) of the Agent Trust Stack.

Quickstart

pip install agent-justice-protocol

from agent_justice_protocol import (
    ForensicInvestigation, IncidentReport, AgentReference,
    EvidenceRecord, JusticeStore,
)

# Investigate an incident
report = IncidentReport(
    reporter=AgentReference(agent_id="agent-monitor-01"),
    incident_type="service_failure",
    severity="high",
    description="Agent failed to deliver contracted output",
    subjects=[AgentReference(agent_id="agent-suspect-42")],
)
investigation = ForensicInvestigation(report)
investigation.add_evidence(EvidenceRecord(
    evidence_type="interaction_log",
    provenance_tier=2,
    source_agent_id="agent-monitor-01",
    content={"action": "request_sent", "status": "timeout"},
))
finding = investigation.produce_finding()

# Store results
store = JusticeStore()
store.append_finding(finding)

Three Modules

Module	Purpose	Key Classes
Forensics Engine	Investigate incidents: collect evidence, reconstruct timelines, flag causal indicators	ForensicInvestigation, IncidentReport, EvidenceRequest, PrivacyGuard
Dispute Resolution	File claims, exchange evidence (commit-reveal), arbitrate with peer panels	Dispute, DisputeClaim, ArbitratorPool, render_decision
Risk Assessment	Score agent risk, compute actuarial outputs for insurance underwriting	RiskEngine, RiskProfile, population_risk_summary

Each module ships independently. Module 1 requires only a provenance chain. Module 2 depends on Module 1 for evidence. Module 3 consumes output from both.

CLI

# Run a forensic investigation
agent-justice investigate --reporter agent-01 --type service_failure \
    --severity high --description "Output not delivered" --subject agent-02

# File a dispute
agent-justice file --claimant agent-01 --respondent agent-02 \
    --finding <finding-id> --harm-type financial \
    --harm-description "Failed delivery" --amount 500

# Check risk profile
agent-justice query agent-02

# Store statistics
agent-justice status

Architecture

Incident
    -> [Module 1: Forensics Engine]
        Evidence collection -> Timeline reconstruction -> Causal indicators
        Output: ForensicFinding
    -> [Module 2: Dispute Resolution]
        Claim filing -> Response -> Evidence exchange -> Arbitration
        Output: ArbitrationDecision
    -> [Module 3: Risk Assessment]
        Historical findings + decisions -> Risk scoring -> Actuarial outputs
        Output: RiskProfile

Key Features

• Evidence-first: Every dispute must be grounded in a forensic investigation
• Provenance-tiered: Evidence weight scales with quality (Tier 1 cryptographic > Tier 4 self-reported)
• Privacy-protected: Anti-fishing rules prevent weaponization of investigations (Rules 5/5a)
• Commit-reveal: Blind evidence exchange prevents strategic adaptation
• Three resolution tiers: Automated (<$1K), peer arbitration ($1K-$50K), human escalation (>$50K)
• Actuarial outputs: Expected loss rate, loss severity distribution, premium basis
• Identity-agnostic: Works with CoC, ERC-8004, A2A, W3C VC/DID, or bare URIs

Optional Integration

pip install agent-justice-protocol[coc]  # Chain of Consciousness integration
pip install agent-justice-protocol[arp]  # Agent Rating Protocol integration

Security Disclaimer (VAM-SEC v1.0)

This software is provided as a reference implementation of the Agent Justice Protocol specification. It is intended for research, development, and integration testing.

This software does not provide legal advice, legal representation, or legally binding arbitration. Dispute decisions rendered by this software are structured data outputs, not legal judgments. For disputes exceeding $50,000 or involving regulatory compliance, human legal counsel is required.

Limitations:

• Automated causal analysis (Phase 4a) produces advisory indicators only — not causal conclusions
• Risk scores are statistical estimates, not guarantees of future behavior
• Actuarial outputs are starting points for underwriters, not insurance quotes
• The commit-reveal protocol assumes honest-majority arbitrator panels

No warranties. See LICENSE for full terms.

License

Apache 2.0 — see LICENSE.

Links

• Whitepaper — Full Agent Justice Protocol specification
• Chain of Consciousness — Provenance chain protocol
• Agent Rating Protocol — Reputation system
• AB Support — Organization homepage