Source: arXiv

Paper

Neuro-Symbolic AI for Korean Criminal Law: Sentencing Prediction and Document Drafting — Yeonseok Lee (submitted 22 Jul 2026)
ID: arXiv:2607.19740

What it claims

Korea’s summary proceedings (guyaksik) move high-volume minor cases — simple DUI, unlicensed driving, minor traffic casualties — quickly, but they dump heavy drafting load on prosecutors. Pure LLMs can extract facts, yet their probabilistic arithmetic is a bad fit for statutory fine math: legal accountability has almost no room for stochastic error.

The proposed neuro-symbolic split: the LLM does semantic extraction only; statutory fine calculation is offloaded to a Satisfiability Modulo Theories (SMT) solver. A human-in-the-loop check keeps professional oversight. The pipeline formalizes the 2026 Sentencing Guidelines for Traffic Offenses as a deterministic support path for summary indictments — not a judge replacement.

The breakdown

Summary traffic cases look “simple” from outside Korea — DUI, unlicensed driving, minor casualties — but volume is the point: guyaksik exists to clear that stack. The paper’s load-bearing claim is architectural. LLMs stay in the language lane (pulling unstructured facts into structured slots). The arithmetic of statutes moves to an SMT solver where outcomes are checkable. Human-in-the-loop is not a slogan here; it is how the design keeps professional accountability when the system drafts toward a summary indictment. Formalizing the 2026 Sentencing Guidelines for Traffic Offenses inside that pipeline is the Korea-specific hook: the statutes are not a toy dataset.

Why readers outside the lab should care

If you drive, insure, or advise people under Korean traffic criminal process, “ChatGPT said my fine would be X” is the wrong tool class. This paper is a blueprint for why Korea-facing legal AI should cage models away from statutory arithmetic. Foreign counsel and compliance teams evaluating Korean legal-tech vendors should ask whether math lives in code/solvers or in the model’s next token.

For compliance readers comparing legal-tech stacks across jurisdictions, the Korea case is a stress test: high volume, tight arithmetic, low tolerance for model improvisation.

What travelers and expats should watch

• Do treat any chatbot “sentencing estimate” as non-authoritative — even when fluent in Korean.
• Do expect future Korean legal tools to advertise extraction + calculator splits; ask vendors where the statute math runs.
• Don’t read this preprint as enacted court software or as legal advice for a real case.
• Expect human review remains part of the proposed design — automation without oversight is explicitly avoided.
• Do keep embassy/legal counsel channels for real cases; a research pipeline is not a substitute for representation.

Context

Read this as architecture for low-tolerance legal arithmetic, not as “AI will decide Korean guilt.” The Korelay frame: Korea’s summary-traffic stack is exactly where probabilistic prose must not own the numbers. If a vendor demo shows a single chat box that both narrates the facts and emits a fine, ask where the solver (or equivalent deterministic layer) lives — and who signs the human check.

Source

arXiv:2607.19740 — abstract and framing cited; open the OA PDF for architecture details and guideline formalization. Not legal advice. Do not republish the PDF.