5.2. Letter of Credit Risk Analysis Engine (LEO)

LEO (Letter of Credit Evaluation through Ontology-driven Inference) is a powerful rule-based detection engine that automatically analyzes the Letter of Credit (L/C)—one of the core payment instruments in global trade—to identify hidden toxic clauses and risks.

1. Overview of LEO's Main Features

Detection and risk analysis

• Deterministic Rule Checker: It checks clear conditions stated in the L/C—such as shipment deadlines, document presentation periods, and who bears bank fees—in real time through a configured rule engine.
• Risk Identification: It detects clauses unfavorable to the exporter (e.g., abnormally short shipment deadlines, unclear document requirements) and reports the clause's Severity to developers/practitioners via the analysis page.
• Interpretive Issue Focus: It picks out ambiguous parts of the L/C text and creates evidence-search cues (Queries) so that deeper analysis tools such as CYGNUS can review them intensively.

Feedback loop and EVA validation

• LC Review Feedback Loop (Clarification): When ambiguous items are found in LEO's analysis results, the system automatically generates a Clarification request to the user. When the user provides additional information, only that item is re-analyzed asynchronously to refine the risk assessment.
• EVA Validation System (Evaluation-Validation-Assertion): This is the 5th layer (L5) that re-validates LEO's AI analysis results. To prevent false positives and increase analysis accuracy, it performs multiple validation stages:
  • Adversarial Review: The AI raises counterarguments to its own analysis results to filter out weak evidence.
  • Retrieve & Judge ✨ NEW: It searches and cross-references evidence documents that contradict the judgment from the actual knowledge graph, and re-evaluates the judgment (implemented 2026-04).
  • Counter-Evidence Search: It searches the knowledge graph for regulations/cases that contradict the analysis result.
  • FP Pattern Detection: It compares against past false-positive case patterns to block similar false positives in advance.

Run options

• Stage 3 toggle ✨ NEW: In the LC analysis modal, you can turn Stage 3 (deep GraphRAG + EVA) on and off. When you need a quick check, turn Stage 3 off to see results within 1–2 minutes; for final review, turn it on to get results with EVA validation completed.
• LEO Agentic Tool ✨ NEW: If you tell Lyra "Analyze the risk of this LC", a multi-stage analysis runs internally via the LEO Agentic Tool route, and results are streamed as real-time socket events. Even if you leave and return to the page mid-run, it recovers to the REVIEWING state.

1-1. LEO Troubleshooting (FAQ)

• Q: "The loaded L/C all passed with no errors, but could problems still arise later?"
  • Error code/symptom: LEO_FALSE_NEGATIVE
  • A: LEO's role is to catch, in the first pass, rules like 'clearly standardized figures or time limits.' Higher-order contextual risks, such as nuances in specific wording, are handled by the secondary deep-analysis engine CYGNUS, so you should cross-check the results of both systems.
• Q: "There are too many items under 'LEO Failures' on the analysis screen."
  • Error code/symptom: LEO_RULE_FAILURES
  • A: A warning (Fail) appears when the L/C's issuance conditions violate common standards (e.g., UCP 600) or are written severely unfavorably to the exporter. Don't ignore them and proceed; send the result report to the customer's issuing bank and request an Amendment to the conditions to block commercial risk.
• Q: "Can I customize (modify) the L/C criteria and rules LEO checks with our company's own strict standards?"
  • Error code/symptom: LEO_CUSTOM_RULESET_UNSUPPORTED
  • A: Currently LEO operates based on global-standard commercial regulations, and analysis customization by injecting a per-group (company) custom ruleset is not yet supported.