lawdangle-kr 0.1.0

Dead cross-reference detector for Korean statutes

lawdangle-kr

Dead Cross-Reference Detector for Korean Statutes Detects and classifies references in in-force Korean laws that point to repealed, renamed, transferred, or obsolete target laws.

한국어 README → README.ko.md

---

Existing citation-checkers verify only whether the citing article itself exists. lawdangle asks a different question: does the cited law still point at a living coordinate? — that is, it catches dangling references to laws that have been repealed, renamed, transferred, or hollowed out.

	Checks	Example
Existence check	Whether the citing article is real	"Criminal Act art. 9999" → does not exist
lawdangle (liveness)	Whether the cited target was repealed / renamed / transferred	"「National Balanced Development Special Act」 art. 17(2)" → repealed & transferred

The direction is reversed: the former catches fakes inside the text; the latter catches dead references inside the text.

The core trap: naive string-replacement of law names produces mis-corrections in split-transfer and article-renumbering cases — sending a wrong reference somewhere more wrong. lawdangle therefore detects and proposes, but never auto-applies corrections.

The five categories

Code	Type	Works?	Remedy	Severity
A	Simple rename	Yes	Replace name only	5 (low)
B	Full amendment / transfer (article mapped)	Yes	Replace name + article	4
C	Split succession (1 → N)	Yes	Manual item-level mapping	2 (high)
D	Obsolete (premise dissolved)	Usually self-contained	Delete / restructure the citation	3
E	Pure repeal (empty reference)	May break	Legislative fix	1 (highest)

Severity ranking (remediation priority): E > C > D > B > A. C / D / E are never auto-asserted — they are flagged for human review with evidence attached. This restraint is the heart of the tool's reliability.

Automatic classification scope (live API)

Verdict	How	Confidence
In-force / Repealed	target=law · eflaw lookup	Auto (high)
A — Rename	Law-ID continuity (old name absent, same ID alive under a new name) + cited-article survival check	Auto (high)
A — Official abbreviation	Official abbreviation dictionary (lsAbrv, ~2,600 entries) → full name	Auto (high)
A — Truncated name	Conservative endswith partial match → full name	Auto (medium)
B — Rename + article moved	Renamed, but the cited article is deleted/renumbered in the current text → suggests --map	Auto candidate (medium)
B — Transfer / C — Split	Repeal + successor candidates (addendum back-trace + amendment-reason) → article mapping is manual	Flag + candidates
D — Obsolete / E — Repeal	Repeal + absorption signal (general account) / no successor	Flag
Informal abbreviation (not in dictionary)	Exact match fails → UNKNOWN (mis-match avoided on purpose)	—

Install

pip install lawdangle-kr

Requires Python 3.10+. Live lookups need a free 법제처(MOLEG) OPEN API key (OC), issued at https://open.law.go.kr.

Quick start

# RECOMMENDED — analyze a whole statute by name (fetches the in-force body itself).
# No copy-paste, no case-law contamination, and the citing article is filled in.
LAW_OC=your_oc lawdangle --law "공유수면 관리 및 매립에 관한 법률" --format csv

# Analyze pasted statute text
LAW_OC=your_oc lawdangle path/to/law.txt --format summary

# Offline demo (history fixtures — no API key needed)
lawdangle examples/sample_corpus.txt \
    --fixture test/fixtures/gongyusumyeon.json test/fixtures/deunggi.json \
    --format csv

# --deep: also map the concrete corresponding article for B / rename+moved cases
LAW_OC=your_oc lawdangle path/to/law.txt --format csv --deep

# Article-correspondence helper (old article → successor article)
# Successor auto-discovered when omitted (works for split transfers too):
LAW_OC=your_oc lawdangle --map "국가균형발전 특별법" "제17조제2항"
# ...or specify the successor explicitly:
LAW_OC=your_oc lawdangle --map "국가균형발전 특별법" "제17조제2항" \
    "지역 산업위기 대응 및 지역경제 회복을 위한 특별법"

from lawdangle import run_law
from lawdangle.resolver import LawGoKrResolver

resolver = LawGoKrResolver("your_oc")
for r in run_law("공유수면 관리 및 매립에 관한 법률", resolver):
    if r.category:                              # skip live/normal references
        print(r.citation.citing_article, r.citation.cited_law_name,
              r.citation.cited_article, "→", r.category.name, "|", r.note)

# Or work at the citation level directly:
from lawdangle import parse_citations, classify
for c in parse_citations(open("law.txt", encoding="utf-8").read()):
    res = classify(c, resolver.resolve(c.cited_law_name))
    print(res.category, res.note)

Pipeline

[statute text]
   ① Parser     extract 「law」 + article citations
   ② Resolver   MOLEG API — current status, rename, successors (cached)
   ③ Classifier 5-way decision tree (+ confidence, manual flag)
   ④ Reporter   CSV / JSON / summary

How successor detection works (live)

The search API has no "successor law" field, but the repealed-law detail (lawService.do) does. lawdangle back-traces successors via two routes:

• Repeal-by-other-law back-trace — when a law was repealed by another law's addendum (타법폐지), the addendum header Addendum(<the repealing law>) … repeals 「self」 names the repealing law = the successor. e.g. 저탄소 녹색성장 기본법 → 「기후위기 대응을 위한 탄소중립ㆍ녹색성장 기본법」; 국토이용관리법 → 「국토의 계획 및 이용에 관한 법률」.
• Amendment-reason prose — extracts 「successor law」 names from the enactment/amendment reason (filtering out the addendum's "amendments to other laws", which are collateral, and keeping only law-shaped names).
• Absorption signal (absorbed) — phrases like "…merged into the general account" mean no successor law → a strong D signal.

Candidate count drives the class: 1 → B (transfer), 2+ → C (split), 0 + absorption → D. All are flagged for manual confirmation.

Content-based successor discovery (split transfers)

The back-traced candidate is the law's general successor — which often differs from the specific provision's successor (the split-transfer trap). For example, 국가균형발전 특별법's general successor is 「지방자치분권 및 지역균형발전에 관한 특별법」, but its art. 17(2) 산업위기대응특별지역 scheme actually moved to a different law: 「지역 산업위기 대응 및 지역경제 회복을 위한 특별법」.

lawdangle finds that automatically: it takes the distinctive concept terms from the old article's text (e.g. 산업위기대응특별지역) and searches current law names for them — the law that defines the scheme surfaces (usually a single hit), and article mapping follows. So --map works even when you do not supply the successor, and --deep attaches the discovered successor + article to B/C results.

Article correspondence (--map, --deep)

For transfers (B), --map narrows "old §17(2) → which successor article?". It walks the old law's history to find the last substantive text before deletion (a cited article may have been deleted before the law itself was repealed), then ranks the successor's current articles by text similarity. When the citation has a paragraph (②), it further narrows down to the paragraph inside the top article (item-level stays manual — that is category C).

$ lawdangle --map "국가균형발전 특별법" "제17조제2항" "지역 산업위기 대응 및 지역경제 회복을 위한 특별법"

옛 법령 : 「국가균형발전 특별법」 제17조  (substantive text from the 2022-01-13 version)
후속법령: 「지역 산업위기 대응 및 지역경제 회복을 위한 특별법」
판정    : manual — multiple close candidates (possible split) (paragraph: §1, sim 0.205)
후보(by similarity):
  1. 제10조제1항  (0.18)   ① the Minister … upon receiving a designation application …
  2. 제12조       (0.176)  …
  3. 제9조        (0.172)  …

It surfaces the design's known answer §17(2) → art. 10 as the #1 candidate (제10조제1항), while not auto-confirming (arts. 9/10/12 are close — possible split) and leaving the final call to a human. That is precisely how the "mis-correction" trap is avoided.

Validated cases (regression fixtures)

Input	Citation	Answer
공유수면법 §13①14	「국가균형발전 특별법」 §17②	B — institution transferred to 「지역 산업위기 대응 및 지역경제 회복을 위한 특별법」
등기특별회계법 §3	「국유재산관리특별회계법」 §6	D — special account abolished & absorbed into the general account; the citing article still works on its own

These two are pinned in test/fixtures/; if the output is not B / D, the regression fails.

Scope (what it deliberately does not do)

• It detects and classifies only. Automatic name-substitution is not a default — corrections are emitted as suggestions; a human applies them. (Auto-substitution causes mis-corrections in B/C cases.)
• It does not equate "repealed" with "defective". D works, B works; only E (and some D) actually breaks enforceability. The report keeps this tone so its findings hold up.
• Informal abbreviations not in the official dictionary resolve to UNKNOWN rather than risk a wrong match.

Development

pip install -e ".[dev]"
pytest                     # offline tests run anywhere; live API tests run only when an OC key is present

Live tests read the OC key from LAW_OC or a local .env (oc=...); without it they are skipped.

See DESIGN.md for the full design rationale.

License

MIT — see LICENSE.