Judicial Dissolution in Chinese Case Law: The Grant Rate Is 61 Per Cent or 25 Per Cent, Depending on What You Count

Ask a legal AI system how often Chinese courts dissolve a company at a shareholder’s request and you will usually get one number. Behind that number is a choice nobody made on purpose: what goes in the denominator. We ran the question against our own corpus today. Count only judgments that decided the claim, and courts ordered dissolution in 60.5 per cent of first-instance cases. Count every case that reached a disposition, and the figure is 25.5 per cent. Same cases, same parser, same day. The difference is that half of all dissolution suits never get a judgment.

This is the third piece in our company-law series, after capital acceleration under three names and shareholder inspection suits. Those measured vocabulary and metadata. This one measures outcomes, which is where most analytics products make their promises, and it shows three places where an outcome statistic quietly changes: the document types you admit, the text you can actually read, and the cases your filter can see.

The claim, briefly

A shareholder holding at least ten per cent of the voting rights may ask a court to dissolve a company whose operation and management are in serious difficulty, where continued existence would cause substantial loss to the shareholders and the difficulty cannot be resolved by other means. The rule was Article 182 of the Company Law as amended in 2018 and is Article 231 of the revised Company Law in force from 1 July 2024. The Supreme People’s Court’s second interpretation of the Company Law fills it in: shareholders’ meetings that cannot be held, or cannot pass resolutions, for two years or more; a board deadlock the shareholders cannot break. It also closes doors. Grievances over information rights, profit distribution or losses are not grounds for dissolution, and courts are told to try mediation first, including buy-outs of the petitioning shareholder.

That last instruction matters for the data. A dispute that ends with one shareholder buying out another rarely ends in a dissolution judgment. It ends with a ruling that the plaintiff has withdrawn.

Finding the cases

We took the union of two routes: the cause-of-action label 公司解散纠纷 (company dissolution dispute), on 15,343 records, and the phrase 公司解散 in the document title, which added 7,731 records the label did not reach. 7,573 of those carry no cause value at all. That is 23,074 records. We then normalised case numbers (full-width and half-width brackets, whitespace) and de-duplicated on case number plus document type, not case number alone, because one case often produces a judgment and a ruling that share a docket. That left 19,631 documents, and the case number’s own code sorted them into first instance (初) and appeal (终).

The first surprise is the document mix. Among those 19,631 documents, 10,233 are rulings and 8,216 are judgments, with 358 mediation statements, 92 notices and 725 whose title names no document type. A dataset described as “company dissolution judgments” is, by count, mostly not judgments.

What happened in 14,001 first-instance cases

We parsed the operative part of every first-instance document, the passage after 判决如下 or 裁定如下, and assigned each distinct case number one final disposition. Where a case had several documents, a decision on the claim outranked a withdrawal, and a withdrawal outranked interim rulings on jurisdiction, preservation or procedure. 375 first-instance case numbers had more than one document.

From one table, three honest answers to “how often do courts dissolve the company?”:

• 60.5 per cent of cases decided on the merits (3,013 of 4,979).
• 25.5 per cent of cases that reached any disposition (3,013 of 11,819).
• 21.5 per cent of all first-instance case numbers we found, if every unreadable or unresolved case is treated as not granted (3,013 of 14,001).

None of these is wrong. They answer different questions. The first tells counsel what happens if the other side fights to judgment. The second is closer to what a client filing a petition should expect. The trouble starts when a product reports one of them without saying which, or when two products report different ones and a buyer compares them.

Examples of each disposition, from the documents themselves: in (2021)沪0151民初3152号 the Shanghai Chongming District People’s Court ordered the dissolution of the defendant company in a single sentence. In (2021)粤0309民初15401号 the Shenzhen Longhua District People’s Court dismissed all of the plaintiff’s claims; that judgment is held three times in our index across two ingest batches, which is why we de-duplicate before counting. In (2021)鲁1092民初434号 the Weihai Economic and Technological Development Zone People’s Court permitted the plaintiff to withdraw, and in (2021)渝0112民初3161号 the Chongqing Yubei District People’s Court treated the suit as withdrawn. Neither withdrawal ruling gives a reason.

How well the parser reads the outcome

Rules that read Chinese operative parts are easy to write and easy to get slightly wrong, so we sampled each class at random. Of 40 cases classed as dissolution ordered, 40 had an operative part ordering dissolution. Of 25 classed as dismissed, 24 dismissed a dissolution claim and one was a labour dispute; of 25 withdrawn, 25 permitted or deemed withdrawal; of 12 struck out, 12 struck out the suit. These samples show the classes are close to clean, not that the parser finds every case that belongs in them.

That labour dispute points to the weakness of the title route rather than the parser. Chinese titles name the parties, and a party can be the liquidation group of a dissolved company, so a lease, wage or heating-supply dispute can carry the words 公司解散 in its title. 93 first-instance case numbers in our set carry only some other cause label for that reason, and a handful of unlabelled ones do the same. Removing the 93 moves the merits grant rate from 60.5 to 60.8 per cent and the share of all dispositions from 25.5 to 25.6 per cent, so we left them in and report it here.

The first version of the grant rule was not clean, and the way it failed is worth knowing. It counted any operative part containing 解散 (dissolve) that did not also dismiss a claim. Two of 25 sampled “grants” were lease disputes brought by the liquidation group of an already dissolved enterprise, whose name contains the word. Titles containing 公司解散 pulled them into the set, and the party name pulled them into the grant class. The rule used here reads each clause of the operative part separately and excludes clauses naming a liquidation group, a dismissal, a refusal to dissolve, a finding that a company is already dissolved, or a criminal conviction. It also catches wordings the first version missed, such as a company “dissolved from the date this judgment takes effect”.

The text cap hid more than one judgment in eight

When we ran this analysis on 15 September 2026, 1,920 of the 23,074 dissolution records had a body of exactly 6,000 characters. Only 6 records in the set were longer, so this was a storage cap, not the length of the judgments. The cap was backfilled later in September 2026; re-measured on 24 September 2026 it affected 0.003 per cent of the corpus, and the counts in this section describe the corpus as it stood on 15 September 2026. A Chinese judgment puts its operative part at the very end, so a capped record keeps the parties, the evidence and often the reasoning, and loses the result. 801 first-instance cases had no readable disposition for this reason, 769 of them judgments. Set against the 4,979 merits judgments we could read, that is more than one judgment in eight without a result.

Truncation is unlikely to be random. The judgments that exceeded 6,000 characters were the long ones, which is where contested evidence and extended reasoning end up; we have not measured whether their outcomes differ. If every capped judgment were a dismissal, the merits grant rate would be 52.1 per cent; if every one were a grant, 66.0 per cent. The case (2021)琼02民初292号, a first-instance judgment of the Sanya Intermediate People’s Court, stopped mid-sentence in the plaintiff’s response to evidence in the copy we read. Any outcome model trained on a capped corpus without knowing about the cap learns from a set that is missing its hardest cases.

The cause label drops cases, but not the rate

As with inspection suits, the cause label misses whole years. By distinct first-instance case number it reaches 8,737 of 14,001 (62.4 per cent). For cases filed in 2013 to 2016 it reaches 81.4 per cent, in 2017 61.6 per cent, in 2018 to 2020 only 15.5 per cent, and from 2022 to 2025 99.6 per cent. The same two ingest batches with no cause values are responsible.

Here, unusually, the gap does not move the headline. Using the label alone, the merits grant rate is 59.8 per cent, the grant share of all dispositions 25.3 per cent and the withdrawal share 52.1 per cent, against 60.5, 25.5 and 50.1 for the union. The missing cases are missing by ingest batch, and batches do not choose outcomes. That is a result worth checking rather than assuming: the label would bias any count by year, and it would bias an outcome rate if one batch held a different mix of courts or document types. Here it does not.

By filing year, the rates move in opposite directions

We bucketed cases by the year in the case number, which is the year the case was docketed. It does not depend on the index’s date field, which in some batches holds a publication date rather than the date a judgment was signed.

Read one column and courts have become far more willing to dissolve companies: 78.8 per cent of merits judgments in cases filed in 2025. Read the next and a petitioner’s odds have fallen, to 18.3 per cent. Both movements are mostly the same thing seen from two sides: withdrawals have grown from under half of dispositions to three in four.

We cannot tell you how much of that is behaviour. Recent filing years are incomplete in two ways that both push in this direction. A withdrawal ends a case in weeks, while a contested judgment can take a year or more, so the 2024 and 2025 cohorts are missing many cases still being fought. And publication of judgments fell sharply from 2023, without any guarantee that it fell evenly across document types. A dashboard that shows the 2025 grant rate as a trend line is showing, at least in part, which cases have finished and been published.

On appeal

We found 3,586 second-instance case numbers. 2,612 had a parseable outcome: the appeal was dismissed and the decision upheld in 1,629 (62.4 per cent), withdrawn in 476 (18.2 per cent), reversed or varied in 359 (13.7 per cent) and remanded in 148 (5.7 per cent). The cap is worse here: for 811 appeal case numbers the text stops before the operative part, because appellate judgments recite the first-instance findings before adding their own. We did not work out which side appealed, so a reversal here can run either way.

How to reproduce this against any corpus

1. Take the union of the cause label and the dispute name in titles, and de-duplicate on normalised case number plus document type.
2. Split by the case number’s own procedure code, not a procedure field, and count the document types before you count outcomes.
3. Parse outcomes clause by clause, then draw a random sample from every class, including the one you expect to be clean.
4. Count records whose body ends at a round number of characters, and report outcome rates with bounds for them.
5. Publish at least two denominators, merits decisions and all dispositions, and bucket years by docket year rather than a date field.

What we would hand a buyer

If a vendor quotes you a win rate for Chinese company dissolution suits, ask four questions. Does the denominator include withdrawals? Were rulings and mediation statements in the set, or only judgments? How many records in the set end at a fixed character count? And are recent years shown against docket year or publication date? A product that cannot answer the second and third questions cannot tell you whether its number is 25 per cent or 60. We can supply the record identifiers and the class assigned to each, so the same comparison can be rerun on your side.

What we are not claiming

We did not measure any other provider’s data. These are rates among published documents in one corpus, not rates among all petitions filed in China, and nothing here predicts the outcome of a particular case. The class samples show that the classes are clean, not that no case was misclassified, and 1,252 first-instance cases have no readable outcome at all. The title route is a floor, since titles that omit the dispute name are not counted, and it also admits about a hundred off-topic cases that name a dissolved company’s liquidation group as a party. We did not read why plaintiffs withdrew, did not code the dissolution grounds courts relied on, and did not separate cases involving the same company. We fixed nothing in the corpus for this analysis: the 6,000-character cap (backfilled corpus-wide later in September 2026), the blank cause values and the duplicate records were all still there on 15 September 2026, documented here instead of silently changed.