Judge and Panel Analytics From Chinese Judgments: What the Signature Block Supports, 2013–2025

Judge analytics is one of the few legal-AI features buyers will pay extra for. In the United States it is built on dockets that name the assigned judge in a structured field. Chinese judgments have no such field. What they have is a signature block at the end of every document: the presiding judge, the other judges, any people's assessors, the date, the judicial assistant and the clerk, each with a title and a name.

That block is enough to answer some questions well. Was the case heard by one judge or by a panel? Did lay assessors sit? How has that changed since China widened single-judge trials? It is not enough to answer others, and the first number most people will compute from it is wrong. We measured both on our own index. Three findings matter for anyone building on this data:

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 shares here are shares of judgments in our index, not of cases filed: many simple cases end in mediation or withdrawal and never produce a judgment. Where a number depends on a choice we made, the choice is stated.

What the signature block contains

A typical civil first-instance judgment ends like this (layout reproduced, names from a published 2025 judgment):

审 判 员　林启新
二〇二五年十月三十日
法官助理　温秋意
书 记 员　陈俞任

One judge (审判员), a date, a judicial assistant (法官助理) and a clerk (书记员). A collegial first-instance panel adds a presiding judge (审判长) and usually one or two people's assessors (人民陪审员). An appeal panel is normally three professional judges. Older documents also carry acting judges (代理审判员). There is no judge number, no court-internal ID, and nothing in the metadata: the court, case number and date are fields; the people are only in the text.

How we measured

• Population. Civil first-instance records are those whose case number contains 民初; civil appeals contain 民终. Judgments are records whose title contains 判决书. One batch (the 2025 ingest, 3.8M records) has titles ending in “一审民事案” with no document type, so there we classified by the operative formula instead: the body says 判决如下 and does not say 裁定如下. Years 2013–2025.
• Exclusions. Two batches whose bodies are almost entirely empty, and a 7,000-record test set. For the pooled annual series we also drop the two batches whose date field is a publication date rather than the decision date (see our date-field audit). One remaining batch, the only source for 2024, has the same problem for about half its records; its row is flagged.
• Detection. A record has a signature block if it contains 书记员 (clerk). It has a collegial panel if it contains 审判长, and assessors if it contains 人民陪审员. Each title is matched two ways: as the indexed token, and as a phrase of its single characters, which is how the spaced form is indexed. Both are exact-match queries run on the full index, not samples.
• Validation. We drew stratified random samples — 50 first-instance and 30 appeal judgments from each of 12 batch-year strata, 960 in all (Elasticsearch random_score, seed 20260930) — and parsed the signature blocks with a regular-expression parser that splits on role titles. Where the parser found a panel, the index-level presiding-judge flag agreed with it in 757 of 757 records.

The collapse that was a typography change

Chinese court documents are often typeset with the two- and three-character titles spread to the width of a four-character one: 审　判　长, 书　记　员. On screen the columns line up. To a search engine they are different words. Our index uses a standard Chinese tokenizer, which indexes 审判长 as a word but indexes 审　判　长 as three unrelated single characters. A term query for the title finds the first and never the second.

Whether a document is spaced depends on the court's template and on how the batch was ingested. In our samples, none of the 80 records drawn from one 2022–2023 batch had a spaced title near the end; in the other modern batches a quarter to nearly half did. One of the two large pre-2019 batches has the whitespace stripped altogether. So the naive signal does not decay smoothly; it drops where the mix of batches changes, which in our index is 2018.

The naive series says China abolished collegial first-instance trials in one year. The corrected series says the share fell by about a third between 2017 and 2019 and kept falling. Both are “measured on 25-plus million judgments”. Only one is about courts. The same mechanism produces false zeros in phrase search generally; we documented it for causes of action and company names.

What the corrected series shows: first instance

Percentages of presiding judge and assessor are of records with a detected signature block. †The 2024 row comes from a batch whose date field is a publication date for about half its records, so part of it belongs to 2023; read it as “between 2023 and 2025”.

Two things stand out. First, assessors sit on most first-instance collegial panels in every year: in 2017, 38.5% of signed judgments had an assessor against 43.7% with a presiding judge. When the collegial share falls, lay participation in written civil judgments falls with it. Second, the decline starts before the law changed. The NPC Standing Committee authorised a two-year pilot of civil procedure reforms in selected courts at the end of 2019, and the amended Civil Procedure Law, in force from 1 January 2022, allows a basic-level court to try a first-instance case under ordinary procedure with a single judge where the basic facts are clear. Those changes fit the 2020–2023 slope, but the 2018–2019 fall predates them, and our data cannot say why. We report the shape; we do not attribute it.

Legal context on this page is a summary for data users, not legal advice. We did not verify which courts took part in the 2020–2021 pilot.

Appeals heard by one judge

Until 2019 a Chinese civil appeal judgment was, in our index, essentially always signed by a three-judge panel: the presiding-judge share is 99.7% to 99.9% in every year from 2013 to 2019, and in 133 parsed appeal samples dated 2015–2019 we found no single-judge signature. The amended law now lets an intermediate court hear by one judge, with both parties' consent, an appeal from a first-instance case decided under simplified procedure, or an appeal against a ruling, where the facts are clear.

In the samples the 2022–2025 records without a presiding judge are single-judge signatures (25 of 25 parsed). We leave 2020 and 2021 out of the table: once publication-date batches are excluded, fewer than 51,000 appeal judgments remain for those years, mostly from one small batch. They should not be read as a trend point in either direction.

Same year, different batch

The batch effect is visible without any chart. For civil first-instance judgments dated 2017, the naive presiding-judge share is 42.0% in one batch and 13.1% in another. Corrected, they are 42.0% and 48.4%: the whole gap was spacing. For 2022, three other batches agree at 8.7% to 9.7%, while a fourth reports 21.0%. That fourth batch is one whose date field is a publication date; when we parsed the signature dates of its records for one cause of action in an earlier audit, every one we could read was signed a year before the date field says. Its “2022” judgments are largely 2021 judgments, and they carry 2021's panel mix. Ingest batches produce trends of their own; panel composition is one more field where they do.

Three layouts one parser has to handle

1. Spaced titles. 审　判　长, 书 记 员, with ideographic or ASCII spaces, sometimes also inside two-character names (万　峰). Normalise whitespace before matching, and remember that the index did not.
2. The whole panel on one line. In one of the two large pre-2019 batches, whitespace was stripped and line breaks inside the block removed: 审判长王晓梅审判员吕中泉审判员孙延义. In our samples, 96 of the 96 multi-member panels from that batch looked like this. Titles are the only separators; a parser that expects one role per line finds nothing, and one that expects a space after the title reads the whole line as a name.
3. Something after the signature. Many courts append the text of the statutes applied (附：相关法律条文) after the signature block, sometimes several pages of it. A parser that reads a fixed tail misses the panel. Reading the last 500 characters, ours found no panel in 149 of 600 first-instance samples (24.8%): 114 because an appendix followed or replaced the tail, 18 because the body was empty or a non-publication notice, 17 for other reasons. In appeals it was 54 of 360 (15.0%), 48 of them appendices. Search backwards from the end for the last block of role titles instead.

None of this is exotic. It is the difference between “we extract the judges” and a parse rate you can put in a data sheet.

Names are strings, not people

Chinese given names are concentrated, and judges are drawn from the same population. The phrase “审判员王伟” or “审判长王伟” (judge or presiding judge Wang Wei) matches 177,282 records in our index across 484 distinct values of the court field; 张伟 matches 130,063 across 419. These counts include longer names beginning with the same two characters and are an upper bound on any one name, but the point stands: a name without a court and a period does not identify a person, and even court-plus-name breaks when a judge moves courts or a court is renamed (court names split into several buckets in our index too).

The roles themselves also change over time, which matters if you model “who signed” as a feature. In our first-instance samples acting judges (代理审判员) appear in 20 of 100 parsed records from 2015 and 2017 and in none of the parsed records from 2020 onward, consistent with the judge-quota reform that retired the title. Judicial assistants appear in none of those 100 records and in between a fifth and two-fifths of parsed records in each stratum from 2020 on. These sample figures show direction; the strata are 50 records each.

What we did not find is systematic anonymisation of judges. Two of 451 parsed first-instance panels and none of 306 appeal panels had a masked judge name; party names in the same documents are routinely masked. Judge names are published on purpose. That makes them usable, and it also makes per-judge scores something to handle with care in any product shown to litigants.

What to do with this in a pipeline

• Match both forms of every title. Normalise whitespace in your own copy; in someone else's index, query the token and the character phrase and take the union.
• Parse from the end, over the whole document. Find the last contiguous run of role titles and names; do not assume it is in the last N characters.
• Emit a parse status per record (parsed, empty body, appendix only, unrecognised) and publish the rate by batch and year. A panel field with an unknown null rate will be read as “single judge”.
• Keep names as strings plus context. Store court, case-number court code, and decision date alongside each name; resolve identities as a separate, reversible step.
• Keep publication-date batches out of annual comparisons, or re-date them from the signature line first.
• Slice evaluation sets by period. A model tested on 2015 judgments will expect a panel and assessors on a first-instance civil judgment about half the time; in 2025 that happens about twice in a hundred.

How to reproduce this on any delivery

# Elasticsearch, body analysed with ik_max_word. Presiding judge, both forms:
{"bool": {"should": [
  {"term": {"body": "审判长"}},
  {"match_phrase": {"body": {"query": "审 判 长", "analyzer": "ik_max_word"}}}
], "minimum_should_match": 1}}
# Same pattern for 书记员 / "书 记 员" (signature present)
# and 人民陪审员 / "人 民 陪 审 员" (assessor present).
# Population: wildcard caseNo "*民初*" or "*民终*", title phrase 判决书,
# year 2013-2025; aggregate by dataset and year before pooling.

Run _analyze on each title in both forms before you trust a count: if the spaced form comes back as single characters, the unspaced query alone is a lower bound.

What this page does not claim

• Shares are of judgments in our index with a detected signature block, not of all cases filed or tried in China; mediated and withdrawn cases are not in the denominator.
• Annual record counts reflect our collection windows, not litigation volume. Compare shares, not counts.
• “Presiding judge present” is our proxy for a collegial panel. In samples it agreed with panel size in all but two of 757 parsed records, but a quoted lower-court panel in an appeal text could in principle trigger it.
• Samples (50 or 30 per stratum) support direction and order of magnitude, not decimals.
• We do not explain the 2018–2019 decline, and the legal summary is informational, not legal advice.