Shareholder Inspection Suits in Chinese Case Law: The Cause Filter Misses 2018–2020, and the Date Field Delays Article 57 by Six Months

A shareholder who wants to see a Chinese company’s books files a right-to-information suit. It is one of the most routine proceedings in PRC company law, it has its own cause-of-action label, and the 2023 revision of the Company Law rewrote the rule it turns on. That makes it a clean test of two things every legal AI team relies on without checking: that the cause label finds the cases, and that the date field dates them. In our corpus the first assumption fails for three whole years, and the second moves a statutory change by six months.

This continues the company-law series that began with capital acceleration under three names. That piece measured vocabulary. This one measures the two metadata fields that sit in front of the vocabulary. Every figure below is an exact count taken on 14 September 2026 against our own judgment index, with total-hit tracking on, unless it is labelled as a sample.

The rule, and what the revision changed

Under the Company Law as amended in 2013 and 2018, Article 33 let a shareholder of a limited liability company inspect and copy the articles of association, shareholders’ meeting minutes, board and supervisory board resolutions and financial reports, and inspect the accounting books after a written request stating a purpose. The company could refuse if it had reasonable grounds to believe the purpose was improper. Before the 2013 renumbering the same rule was Article 34. The statute said books, not vouchers, and whether accounting vouchers (the entries and the source documents behind them) were within reach became the recurring fight. Some courts reasoned from the Accounting Law that books are compiled from vouchers and ordered both. Others held that the statute stopped at the books.

The revised Company Law, in force from 1 July 2024, answers that question in Article 57: a shareholder may request to inspect the accounting books and accounting vouchers. It also allows a shareholder to engage an accounting firm, law firm or similar intermediary to conduct the inspection, and extends the right to wholly owned subsidiaries. For data purposes the change is useful because it leaves a fingerprint. Judgments that cite the inspection article by number tell you which statute the court applied.

The label finds 65 per cent of the cases

We counted two routes to the same population: the cause-of-action label 股东知情权纠纷 (shareholder right-to-information dispute), and the phrase 股东知情权 in the title of the document. Chinese judgment titles follow a fixed pattern that names the parties and the dispute, so a title match is a strong signal. Body matches are not: of 41,051 records that mention the phrase anywhere, many are share-transfer, dissolution or contract cases that refer to an earlier inspection suit.

Distinct case numbers are counted after normalising full-width and half-width brackets and removing whitespace, which merged 1,019 labelled and 7 unlabelled variants. By case number the label reaches 16,323 of 25,115, or 65.0 per cent. By record it reaches 64.7 per cent. We drew 25 unlabelled title matches at random (seed 20260914): all 25 titles name a shareholder inspection suit, and they include judgments, rulings, a mediation statement and a sanction decision. Two of the 25 have bodies of 8 and 19 characters, so a title search also surfaces records that are nothing but a title.

The gap is not spread evenly. It is almost all in three years.

The cause is ingestion, not law. 10,218 of the 10,722 unlabelled records come from two ingest batches in which no record carries any cause value: one covers judgments dated up to 2019, the other 2020 and 2021. Plot inspection litigation by year with the label as your filter and 2020 falls to 428 records, fewer than any year since 2013. Plot it with the title route and 2020 is one of the busiest years in the series. We have written before about time series that trace ingest schedules. This is the metadata version of it. A column that is blank for a whole batch does not show up as noise. It shows up as a trend.

The date field is not always the judgment date

Chinese judgments end with the date the judgment was signed, written in Chinese numerals: 二〇二一年八月十日, for example. We parsed that line in every labelled record whose date field is 2022 or later. Of 8,962 records, 7,802 gave a signature date we could read. The other 1,160 are truncated bodies, rulings without the usual closing, or dates we could not parse, and are left out.

In batch A the date values cluster on month-end days: 69 records dated 31 January 2024, 59 dated 30 January, 37 dated 29 February. Courts do not sign judgments that way. Collection and publication processes do. One example: (2021)粤1302民初13513号 (Huizhou Huicheng District People’s Court) carries a date field of 26 April 2024 and was signed on 10 August 2021. In most batches the field is a good proxy for the judgment date. In three it is a proxy for when the document appeared, and nothing in the schema says which is which.

Dated correctly, the courts switched to Article 57 within a year

To see what that does to a real question, we matched citations of the inspection article with the statute named in front of them, as in 公司法》第五十七条 for Article 57 and 公司法》第三十三条 or 第三十四条 for the old articles, allowing for a bracketed revision year. Then we bucketed the same labelled records two ways.

By the date field, courts were slow: barely one citation in six named the new article in the half-year it took effect. By the signature date, it was more than half. The lagging batches file judgments signed in 2023 and early 2024 under later dates, where old citations dilute the new ones; we did not decompose the gap batch by batch. From 2025 onward the two measures agree, because the batches that hold 2025 judgments date them correctly. The 13 records signed in the first half of 2024 that already cite Article 57 are plausible, since the revision was promulgated on 29 December 2023, but we have not read them. The old articles do not vanish: 20 of the 75 citing records signed in 2025 H2 still name Article 33 or 34, consistent with courts applying the prior law to earlier facts.

The same error sat in our own previous article. Its half-year table of capital-acceleration wording was bucketed by the date field. Re-counted by signature date, the share reachable only through the statutory wording in 2024 H2 is 35.8 per cent, not 15.8. We have added a correction to that page.

Vouchers became statutory, and mentions went down

If the revision settled the voucher question, you might expect vouchers to be argued less. The mention share among labelled records fell by either measure: by signature date from 48.4 per cent in 2022 H1 to 35.0 per cent in 2024 H2 and 35.8 per cent in 2025 H2, by date field from 55.4 per cent to 43.6 per cent. We did not measure why. A mention can be a claim, a refusal, a quotation of the statute or a list in the operative order, and the fall had begun before July 2024.

The text does show both sides of the pre-revision split, and the post-revision rule being applied. In (2021)兵08民终388号, signed 7 May 2021, the intermediate court of the Eighth Division of the Xinjiang Production and Construction Corps varied a first-instance judgment to add bookkeeping vouchers and source vouchers to what the shareholder could inspect. In (2021)苏09民终4722号, signed 27 October 2021, the Yancheng Intermediate People’s Court ordered financial reports and account books, and its operative order lists no vouchers. In (2025)豫0194民初20349号, signed 28 October 2025, a Zhengzhou court quotes Article 57 and orders both the account books and the accounting vouchers produced for inspection. These illustrate the rule. They say nothing about how often vouchers were granted.

Three zeros that are not zeros

The obvious searches for this doctrine fail silently in our index, for the same reason we described in the acceleration piece: the body field is indexed with one Chinese analyzer and queried with another, and some strings tokenise into positions that never align. The fix we tested today is to analyse the query with the index-time analyzer.

The override recovers the false zeros with no false positive in the samples. It is not free: on the two control phrases it returns 1 and 17 fewer records than the default, so it trades a silent zero for a small silent loss. Across the whole corpus 不正当目的 goes from 0 to 16,904 and 第五十七条, which also matches Article 57 of every other statute, from 0 to 859,665. A retrieval test that asks for “improper purpose” refusals and gets nothing back is measuring the analyzer, not the case law.

How to reproduce this against any corpus

1. Count the cause label, then count the dispute name in document titles, and take the union by normalised case number. Report the share the label reaches, by year.
2. For every year where the label’s share drops, break the unlabelled records down by ingest batch and check whether the cause column is blank for the whole batch.
3. Parse the signature date from the end of each judgment and compare it with the date field, by batch. Look for dates piled on month ends.
4. Re-run any before-and-after comparison around a statutory change with both dates, and publish both.
5. Test every phrase you rely on with both analyzers. Treat any zero as unproven until the index-time analyzer also returns zero.

What we would hand a buyer

If your system answers questions about PRC company-law litigation, ask it for shareholder inspection cases decided in 2020, then run the title route above against the same corpus. A system that routes by cause label will return about one record in eight. Then ask when courts adopted Article 57. A system that trusts the date field will put the majority switch in the first half of 2025 instead of the second half of 2024. Both are metadata errors, and neither shows up in a relevance evaluation that samples from the labelled set. We can supply the record identifiers behind every figure here, including the batch-level date comparison, so the test runs on identical records.

What we are not claiming

We did not measure any other provider’s data, so nothing here is comparative. The title route is a floor: records whose title omits the dispute name are uncounted, and the 25-record sample shows the unlabelled title matches are on topic, not a precision rate. The year column is the index’s own year field, which may share the date field’s lag in the affected batches. Signature dates were parsed for 7,802 of 8,962 records, and the article-citation table counts only records that name the statute before the article number. Mentions of vouchers are not grants. We counted no outcomes and fixed nothing in the corpus: the blank cause columns, the date field and the analyzer mismatch are all still there, documented here instead of silently changed.