KD-Lawset-KO

Korean statute questions in a typed-decision format (choice / noul = yes-no / score = integer), generated by code from public statute text, plus the per-question probability outputs of ThakiCloud/kd-4b-ko-v0 (kotdaihubv2num2-s1). It was built to measure one thing: if a model may abstain, what error rate can we guarantee on the questions it does answer, and how many questions does that guarantee cost?

What is inside

config	split	n	laws	question templates
lawset-v2	calibration / test / shift	532 / 542 / 393	320 / 333 / 245	T1 fine cap exceeded? (noul) · T2 deadline met? (noul, date arithmetic) · T3 fine cap of paragraph k (choice) · T4 prison-term cap in years (score)
lawset-v3	calibration / test / shift	302 / 294 / 158	248 / 236 / 130	HC1 administrative-fine cap of paragraph k (choice) · HS1 how many paragraphs have a fine cap ≥ X (score) · HS2 deadline in days (score, 1–180)
decree-e2	test	591	591	T2 from Enforcement Decrees (시행령) — a set never used for calibration or thresholds

• shift = laws from ministries held out of calibration/test (domain shift).
• Independence: at most one question per (decision type, clause hash) globally, then one per law × decision type.
• Each line: id, template, domain (ministry), law, state (statute excerpt + facts), questions[0] (type, question, options/min/max), gold.
• predictions/*.rows.json: kd-4b outputs (p = probability over options / rungs, label = gold index), same order as the data.
• *.manifest.json: source commit, generator / oracle / converter hashes, drop counters.

How the labels were made — and checked

Labels are computed by code from the statute text (money and date arithmetic, Korean holiday calendar for 2026, 초일 불산입). Every item was accepted only if two independent programs agreed: the generator's own oracle and a separately written, regex-free recomputation. Neither is a human.

Independent LLM audit (2026-09-26): every item was re-solved blind by gpt-5.6-luna (a different model family from the generator and from kd-4b); each disagreement was adjudicated by gpt-5.6-sol. This is not a human audit.

set	items	luna disagreed	sol: code label correct	sol: label wrong	ambiguous	label-error 95% upper bound
lawset-v2	1,467	12	11	0	1	0.25%
lawset-v3	754	0	–	0	0	0.49%
decree-e2	591	13	13	0	0	0.62%
all	2,812	25	24	0	1	0.13%

As an anchoring check, the 25 disagreements were also re-solved by gpt-5.6-sol without the code label: 25/25 matched the code label. The one ambiguous item (t1-ef05045ce360) states a fine as a multiple of an unstated compensation amount, so "is the cap exceeded?" has no single answer.

Correction (2026-09-26, rule scan after the audit). Scanning the line each question targets for multiplier or "whichever is higher" caps (N배, 높은 금액, 초과 금액) found 4 such items, all in lawset-v2 (3 calibration, 1 shift): t1-777a210262b4, t1-9c41a700acfc, t3-d06216b43e5e, t1-ef05045ce360. The LLM audit flagged only the last one; for the other three it accepted the code label, which reads only the fixed amount. Treat all four as ambiguous (listed in flags/variable_cap.json); the released files are unchanged so their hashes stay valid. Removing them moves the lawset-v2 α = 0.10 result from 18.1% abstained / 2.5% error to 19.0% / 1.8%; lawset-v3 and decree-e2 contain none. The generator now excludes these forms for future versions. So the audit bound above is about wrong labels; the ambiguous rate is 4 / 2,812. Upper bounds are two-sided 95% Clopper-Pearson over decided items. Cost: about 1.01M input / 0.14M output tokens on luna and ~20k tokens on sol.

Reference result (kd-4b-ko-v0, Learn-then-Test, δ = 0.10, per decision type)

On lawset-v2 test, targeting α = 0.10 (answered-question error ≤ 10% with probability ≥ 90%), the 4B model answers 81.9% of questions and makes errors on 2.5% of those; α = 0.05 costs 59.6% abstention. The abstained deadline questions of decree-e2 (406 of 591) were solved at 99.3% by an in-house Qwen3.8-27B that only extracted the day count and dates while code did the arithmetic, versus 58.4% for the 4B on the same questions. Thresholds and a small wrapper are in the model repo under selective/.

Limits

• Synthetic, template-generated questions over real statute text. Four + three templates do not cover legal reasoning in general.
• In lawset-v2 and v3 the 4B answers all choice questions correctly (ceiling), so the choice curve carries no information.
• Statute text is the 2026 snapshot of legalize-kr; later amendments are not tracked.

Source and license

Statute text: legalize-kr/legalize-kr (National Law Information Center text; Korean statutes are not subject to copyright, Copyright Act Art. 7). Questions, labels and predictions: CC-BY-4.0, ThakiCloud. No AI Hub data is included. The typed-decision format and serving code come from Kev (Apache-2.0) by Jared Palmer.