China's Civil Code Cutover in Case Law Data: Measuring Statute Citation Drift, 2020–2025

On 1 January 2021 the PRC Civil Code took effect and repealed nine statutes at once, among them the Contract Law, the Tort Liability Law, the Property Law, the Guarantee Law, the Marriage Law and the General Principles of Civil Law. For anyone training or grounding a model on Chinese judgments, that date is the single largest discontinuity in the corpus: the article numbers change, the statute names change, and a model that learned civil law from 2014–2020 judgments will confidently cite provisions that no longer exist.

What is less obvious is that the courts did not switch on that date. Judgments kept citing the repealed statutes for years, and they were right to. We measured how long, in which kinds of case, and what else the cutover left behind in the text. Three findings matter for data work, and only the first was expected.

Everything below describes our own index of 160,291,678 PRC judgment records on the stated date. It is not a census of Chinese court judgments and should not be read as a share of all judgments in China; nobody can state the size of that universe. Where a number depends on a choice we made, the choice is stated so you can rerun it differently.

How the drift was measured

We counted records whose body text contains the full official title of each statute: 中华人民共和国合同法 (Contract Law), 侵权责任法 (Tort Liability Law), 物权法 (Property Law), 担保法 (Guarantee Law), 婚姻法 (Marriage Law), 民法总则 (General Provisions of Civil Law), 民法通则 (General Principles of Civil Law), and 中华人民共和国民法典 (the Civil Code). Each is a phrase query forced through the same ik_max_word analyzer the index was built with. We checked the token sequence for every title first; this matters because the default search analyzer on the same field returns false zeros for many Chinese phrases, a failure we documented in an earlier audit.

Three restrictions shape the numbers:

• Only batches whose date field is the judgment date. In three of our ingest batches the date field holds a publication or collection date, not the signature date; we excluded them, which removes the 2024 row entirely. The date-semantics audit explains how we established which batches are affected.
• Full titles only. A citation written as 《合同法》 without the 中华人民共和国 prefix is not counted. We checked what that costs: in a random sample of 400 civil first-instance records from 2022 (seed 20260929), 167 cited one of these statutes by full title and only 3 more cited one by short title alone. The full-title count captures about 98% of statute-citing records.
• “Cites” means the title appears anywhere in the body. That includes party submissions and, in appeals, the lower court's reasoning quoted back. It is a measure of what the text contains, which is what a model trains on, not a measure of which rule decided the case.

Precision was checked by reading: 60 random 2022 records matched by the Contract Law query all contain the literal title (60/60), none only inside the title of an interpretation.

The cutover month by month

Among records that cite any of the eight statutes by full title, the chart splits each month into three groups: those citing only repealed statutes, those citing both old and new, and those citing only the Civil Code.

The switch was fast in one sense and slow in another. Before 2021 the Civil Code appears in essentially nothing (0.05% of statute-citing records in 2020, mostly discussion of the promulgated but not yet effective text). In January 2021 — the first month the Code was in force — 43.5% of statute-citing records already cite it. But only 30.6% cite it exclusively, and the share still citing nothing but repealed statutes is 56.5%. That old-only group then shrinks steadily: 42.7% by June 2021, 27.5% by December 2021, 17.0% by June 2022 and 11.0% by December 2022.

The record counts per year differ because our collection windows differ by year, not because litigation rose and fell that way; the shares are the comparable figures.

Why the courts kept citing repealed law

This is not stale drafting. The Supreme People's Court's rules on the temporal effect of the Civil Code (最高人民法院关于适用《中华人民共和国民法典》时间效力的若干规定, issued 29 December 2020) provide, as a general rule, that disputes arising from facts that occurred before 1 January 2021 are governed by the law in force at the time. A loan signed in 2019 and litigated in 2023 is decided under the Contract Law, and a well-drafted judgment says so.

The text bears this out. We drew 60 random records from 2023 and 60 from 2025 that cite a repealed statute and do not cite the Civil Code by full title. In both years, 43 of 60 explicitly cite the time-effect rules as the reason for applying the old statute, and 59 of 60 in each year mention a date before 2021 in the body, typically when the contract was signed or the loan made. In none of the 120 does the document's last written date fall in an earlier year than the date field, so these are not misdated 2020 judgments. The old-law tail is judgments doing exactly what the transition rules require.

The time-effect rules themselves are a useful marker. Per hundred statute-citing records they appear 39.5 times in 2021, 37.7 in 2022, 27.5 in 2023 and 10.0 in 2025. A record that cites them is a record in which the court made an explicit choice of regime, which is exactly the label a data team would otherwise have to infer.

Contract law decays slowly, tort law almost at once

The repealed statutes did not fade at the same rate. The table gives each one's share of statute-citing records, so the columns are comparable across years.

The ordering follows how long the underlying facts take to reach a judgment. Contracts and guarantees are signed years before they fail, so contract disputes from pre-2021 agreements were still being decided in 2025. Torts are events, typically litigated within a year or two, so the Tort Liability Law had all but vanished by 2023. The Marriage Law's early drop in 2021 fits the same logic, though family cases are also the category where our index has the most empty bodies in 2017–2021, so we would not lean on its level.

For a model this means the stale-law problem is concentrated where commercial clients care most. A contract-dispute retrieval system in 2023 was still serving judgments in which roughly one in six statute-citing records relied on the Contract Law, with article numbers that map to different Civil Code articles (Contract Law article 107 on liability for breach corresponds to Civil Code article 577; article 114 on liquidated damages to article 585).

Appeals lag first instance by about a year

Splitting by level of court, using all three procedure encodings in our index (the Chinese labels plus two numeric code sets; relying on the Chinese labels alone drops about 30% of first-instance records):

Part of the gap is mechanical: appeals are decided later in a dispute's life, so their facts are older. Part of it is textual: a second-instance judgment usually restates the first-instance reasoning, including the statutes it cited. The 2021 appellate row rests on a thin slice of our index for that year and should be read as direction only.

The statute names that vanish

While reading the old-law samples we kept meeting a citation that cannot be right: 《最高人民法院关于适用时间效力的若干规定》 — “the SPC's provisions on the temporal effect of applying” … nothing. The statute name that belongs in the inner brackets, 〈中华人民共和国民法典〉, is simply not in the text.

It is not rare. 4,056,401 records cite the time-effect rules in some form; 1,837,085 of them (45.3%) contain only the stripped form and another 121,633 contain both. The same loss affects every SPC interpretation whose title nests a statute name. The literal string 关于适用的解释 (“interpretation on the application of”, with no statute) occurs in 14,238,070 records. For comparison, the intact title of the Civil Procedure Law interpretation appears in 17,577,706. We read 200 random records from the stripped group: all 200 contain the literal stripped string, and only 12 also carry an intact nested title elsewhere in the text.

The loss runs through every ingest batch we hold, including the oldest (records dated up to 2019), so it is not a defect of one collection run. We have not established where it happens. The intact citations use four different inner bracket styles — 〈〉 in 225 of 300 sampled, 《》 in 20, ASCII <> in 32, full-width ＜＞ in 13 — and ASCII angle brackets are the kind of thing markup processing removes, but some ASCII-bracketed titles survive intact, so that explanation is at best partial. We report the loss; we do not claim to know its cause.

The consequences are practical:

• Citation extraction. A parser that keys on the statute name will attribute “《最高人民法院关于适用的解释》第九十条” to nothing. The Civil Procedure Law interpretation, the Contract Law interpretations, the Guarantee Law interpretation and the Civil Code interpretations all collapse into the same string.
• Citation graphs and authority scores. Counts of how often an interpretation is cited are understated by a margin that differs by interpretation.
• Regime tagging. A record that cites only the stripped time-effect rules never mentions the Civil Code by name, so a naive “cites 民法典” rule classifies it as old-law only. The unique tail of the title (时间效力的若干规定) still identifies it, and the article number usually disambiguates the rest.

Judgments dated before the rules they cite

The time-effect rules were issued on 29 December 2020. Our index nevertheless holds 666 records dated 2020 that cite them, and the distribution is lopsided: 134 in January, 124 in February, 139 in March, 113 in April, then a thin scatter through the rest of the year. That is the shape of a new-year typo — a judgment signed in early 2021 with the previous year's number.

In a sample of 60 of these records, the last written date inside the document itself also reads 2020 in all 60. So the error is not only in our date field; it is in the judgment text as published, and a signature-date parser will faithfully reproduce it. It is a small number against a corpus this size, but it is the kind that produces confident nonsense in before/after studies: a handful of “2020 judgments applying the Civil Code” is enough to suggest an early-adoption effect that never happened. Any analysis that hinges on a legal effective date should test whether records dated just before it cite rules that did not yet exist.

What to do with this in a data pipeline

1. Tag the governing regime per record, not per year. Minimum signals: full-title and short-title statute citations, citation of the time-effect rules (including the stripped form), and the dates of the underlying facts where they can be parsed. Records dated 2021–2023 are genuinely mixed.
2. Keep old-law judgments; label them. They are correct decisions about disputes that still exist, and contract claims from pre-2021 agreements will keep arriving for years. Dropping them biases a model against exactly those disputes.
3. Carry an article crosswalk. When retrieval or generation surfaces a repealed provision, map it to its Civil Code counterpart and show both. Article numbers do not translate one to one.
4. Repair or flag stripped nested titles. Resolve 关于适用的解释 and its relatives using the article number and surrounding citations, or flag them as unresolved; do not let them silently drop out of citation statistics.
5. Evaluate on time-sliced test sets. A single mixed test set hides whether a model cites the right regime for the right facts. Build separate sets for pre-2021 facts decided after 2021.
6. Plausibility-check dates against legal events. Records dated before the issue date of a rule they cite are cheap to find and should be quarantined from any time-series work.

How to reproduce this on any delivery

Every figure above comes from phrase counts and seeded samples a buyer can rerun on a delivered dataset. In Elasticsearch, force the analyzer the field was indexed with and inspect the token sequence of each title before trusting a count:

POST judgments/_analyze
{ "analyzer": "ik_max_word", "text": "中华人民共和国合同法" }

POST judgments/_count
{ "query": { "bool": {
  "filter": [
    { "match_phrase": { "body": { "query": "中华人民共和国合同法", "analyzer": "ik_max_word" } } },
    { "term": { "year": 2023 } } ],
  "must_not": [
    { "match_phrase": { "body": { "query": "中华人民共和国民法典", "analyzer": "ik_max_word" } } } ] } } }

On a flat file, the equivalent is a literal substring test on the body, which is what our read-back checks used. Samples were drawn with random_score, seed 20260929, field _seq_no; sample sizes are stated beside each result.

What this page does not claim

• It measures our index on 29 September 2026. It is not a statement about Chinese courts as a whole or about any other provider's data.
• “Cites” means the title string occurs in the body. It does not identify the rule that decided the case, and appeal figures include quoted lower-court reasoning.
• The 2021 appellate figure rests on a thin slice; 2024 is excluded entirely; yearly record counts reflect collection windows, not litigation volume.
• The reading samples (60 per year for the old-law tail, 200 for stripped titles, 60 for the 2020 records) support the direction and rough size of each finding, not a precise rate.
• The cause of the stripped nested titles is unknown to us.
• The description of the time-effect rules is a summary for data work. It is informational, not legal advice.