What the Official Chinese Case Database Actually Holds: 5,536 Cases, and Why That Is the Right Number

Somewhere in the first week of a China-coverage project, someone asks the reasonable question: is there an official source? A government database, run by the courts themselves, that we could just use.

There is. The Supreme People's Court operates a case repository, and it is increasingly what a search engine hands back when a data buyer asks where Chinese case law comes from. We checked that this morning: on the buyer-intent query Chinese court judgments dataset English structured API license commercial, the SPC's English page for that database sits in the first half of the first page of results, above several of the academic datasets that used to own the answer.

So we spent this morning reading the thing itself rather than reading about it, and writing down what it holds. The short answer is 5,536 cases. The longer answer is that 5,536 is not a disappointing number, it is the correct number for what that database is — and that the mistake worth avoiding is not using it, but assuming it can do the job a bulk judgment corpus does.

What was measured, and how to re-run it

Three surfaces, all read directly in a browser on 27 August 2026, no intermediary and no search-engine summary:

• The Chinese database at rmfyalk.court.gov.cn — front page, section structure, advanced-search field list, and the statistics call the page makes on load.
• The English landing page for it on the SPC's English site, read as served HTML rather than as rendered impression.
• The English Typical Cases list, paginated to the end, with every entry date recorded.

Anyone can repeat all three in about fifteen minutes. That matters more than the numbers, because official sites change and a buyer who has re-run the check owns the answer instead of citing us.

The Chinese database, read directly

The repository presents a single headline figure on its front page: 5,536 cases. That is not an inference from paginating a list; the page loads a statistics endpoint that returns the same count as a plain integer, stamped with the current month.

The collection is structured in two tiers — guiding cases and reference cases — and divided into five subject sections: criminal, civil, administrative, state compensation, and enforcement. The advanced search is more informative than the case count, because it tells you what the court itself considers the addressable dimensions of a case:

Two things follow from that table. First, the official taxonomy is the same one a bulk corpus has to carry, which is why field normalisation is a shared problem rather than a vendor-specific defect — anyone working with PRC case data ends up mapping the same six dimensions. Second, the interface is Chinese only. Not primarily Chinese, not Chinese with an English toggle: the language switch on the SPC's English site points back at the Chinese main site, and the database itself offers no English mode we could find.

One technical observation, offered as an observation rather than a criticism: the site is a client-rendered application. The served HTML for the front page carries the furniture and almost none of the substance; the case text arrives afterwards via script. There is no robots.txt at the expected path — the request returns a server error rather than a set of rules, which means the crawl posture is undeclared rather than restrictive. For a team wondering why answer engines paraphrase Chinese case law from secondary commentary rather than quoting the official repository, that combination is a large part of the mechanism.

The English channel, counted to the end

Here is the finding we did not expect. The SPC's English site has a page titled People's Court Case Database. It is the page currently ranking on the buyer query. It contains no cases.

Read as served HTML it is roughly sixteen kilobytes, of which the entire body text is 707 characters: a title, a date, a navigation menu, a list of links to other government sites, and a copyright line. There is no case list, no search box, no export, no link to a dataset. It is a signpost with nothing behind it in English.

The English case content lives elsewhere, in a list called Typical Cases, and we paginated it to the end. It holds exactly 100 dated entries across ten pages, running from 17 February 2020 to 20 March 2026. The distribution over those six years is the part worth writing down:

We do not know why the channel behaves this way and are not going to guess. What matters operationally is the shape: publication in English is episodic, so the date on an entry tells you when a translation was posted and nothing about when the case was decided or how current the collection is. The point is easy to confirm from the entries themselves. In the February 2026 burst, guiding cases appear in the order 257, 215, 214, 177, 236, 235, 233, 232 — a translation queue being drained, not a chronology.

Three further details a buyer should know before treating this list as a dataset. Some entries are batch releases: a single line reading (First Batch) can contain many cases, so 100 entries is not 100 cases and the true case count cannot be read off the list. Party names are redacted in the translations, which is appropriate for publication and fatal for any task requiring entity resolution. And the redaction markers themselves are not normalised — on a single page we counted [redacted], [Redacted] and [REDACTED], alongside a mix of full-width and half-width parentheses. That is a hand-edited editorial channel, and it should be read as one.

Why 5,536 is the right number

It would be easy to write the previous two sections as a complaint. It would also be wrong. A curated authority set is supposed to be small, and the smallness is the product.

China's system does not run on binding precedent in the common-law sense, but it does run on designated guidance: a small, deliberately chosen body of cases that lower courts are expected to consult when deciding like matters. The value of that body comes from its selectivity. A guiding case has been argued over, edited, given a formal number and published because someone decided this is how the rule should be read. Add a million routine judgments and you have destroyed the signal you were curating for.

So the official database is complete at 5,536, in the same way a statute book is complete. The failure mode is not that the number is small. It is that a curated authority set and a bulk judgment record are different artefacts, and a scoping exercise that treats one as a sample of the other will produce a plan that cannot be built.

The pyramid the curated set sits on top of

The clearest way to see the difference is by court level. Our index carries the deciding court on nearly every record, so the distribution is directly measurable:

Curation samples the top of that pyramid, because doctrinal significance concentrates there. Retrieval lives at the bottom, because that is where the disputes a legal AI product will actually be asked about get decided. A landlord fee dispute in a county court will never be a guiding case, and it is exactly the kind of matter a user will paste into a search box expecting comparable outcomes.

The weighting inverts as well, and the inversion is sharp. The official database gives state compensation one of its five sections — a fifth of the navigation. In our bulk index, records tagged to state compensation and judicial relief total roughly 257,000, about 0.16% of the corpus. Enforcement runs the other way: also one section out of five in the curated set, and roughly 23.6% of the bulk record. Neither weighting is a mistake. Curation weights by importance; a docket weights by volume. Reading section counts as coverage will mislead you in both directions at once.

One more curve worth seeing before you plan around the top tier

If you decided to build on high-authority material only, the natural move is to filter a bulk corpus down to Supreme People's Court decisions and work with those. Here is what that stratum looks like in our index, by year:

Three years — 2018 through 2020 — account for 765,320 records, which is 91.7% of every Supreme People's Court record in the index. The years on either side are thin by two orders of magnitude. This is our corpus's view of what reached the public record and how much of it we hold; it is not a statement about how many cases the court decided, and we have not established the cause. But if your architecture assumes a steady annual supply of apex-court material, that assumption should be tested against real counts before it reaches a roadmap, and the same discipline applies to recency across the corpus as a whole.

Which artefact for which job

The practical output of all of the above is a routing table. Most production systems need both sources, held separately, never averaged together.

Read that table alongside the academic datasets a buyer typically finds first, and the landscape resolves into three distinct things that are constantly confused for one another: an official curated authority set, frozen research benchmarks, and a bulk record of what courts actually decided. Each answers a question the other two cannot.

The one place the two artefacts should touch

Keeping the sources separate does not mean keeping them apart. There is exactly one workflow where a curated authority set earns its keep inside a bulk-corpus system, and it is worth naming because teams reliably miss it.

Use the curated set as the evaluation set, not as the index.

The hardest problem in Chinese similar-case retrieval is that you cannot tell a good result from a plausible one without a ground truth, and building ground truth by hand is the expense that quietly kills pilots. A guiding case comes with something close to a free label: the court has already stated which rule the case turns on and how that rule should be read. That makes it a usable probe. Take the fact pattern from a guiding case, strip the identifying language, run it through your retrieval stack over the bulk index, and ask whether the cases that come back are governed by the same rule. If they are not, you have found a real defect — not a stylistic one — and you have found it with a few hundred hand-checked probes rather than a few thousand.

Two cautions come with that. The probe set is small and skewed toward doctrinally interesting matters, so a system that scores well on it has not been shown to work on routine disputes; it has been shown not to fail on hard ones. And the guiding cases must stay out of the index you are evaluating, or you are measuring memorisation. This is the same discipline as any held-out set, and it is worth writing into a pilot agreement, because the question of who owns the evaluation artefacts you build during a trial is one of the terms buyers most often leave undefined until it matters.

What this page cannot tell you

• It is one reading, on one date. Everything above was observed on 27 August 2026. Official sites change without notice, and the correct use of this page is as a method to re-run, not a figure to cite indefinitely.
• We did not open a case record end to end. We read the repository's structure, statistics and search surface. We did not verify what a full case page contains for an anonymous visitor, and we make no claim about registration, quotas or export.
• The absent robots.txt is not a prohibition. A server error at that path means no rules were served. It is not permission and it is not a block, and nothing on this page speaks to the lawfulness of any particular acquisition method — a separate question with its own answer.
• 100 entries is not 100 cases. Batch releases make the English list's true case count unknown from the outside, and we have not tried to resolve it.
• Our own figures carry known defects. The court-level counts above are drawn from an index whose court field we have publicly documented as containing thousands of whitespace and entity variants; roughly 9,800 apex-court records sit under variant keys and are excluded from the 834,967 figure. Nothing here should be read as a claim that our normalisation problems are solved. They are not.
• No inference about other suppliers. We have measured our own index and two public government surfaces. We have not measured anyone else's corpus and cannot say how these patterns compare.

The short version

China has an official case database, it is real, it is well built for its purpose, and it holds 5,536 cases. Its English presence is a landing page with no cases on it and a hand-curated list of 100 entries accumulated over six years, seven in ten of which arrived either in 2020 or in a five-week burst this spring.

None of that is a scandal. It is what an authority collection looks like, and what a translation programme looks like. The scandal would be a scoping deck that read official database and wrote corpus, then discovered in month four that similar-case retrieval has nothing to retrieve from and every base rate in the product measures an editorial committee.

The honest framing is that these are complements. If you need to know what the Supreme People's Court has designated as the correct reading of a rule, go to the curated set; nothing we hold can substitute for it, and we would tell a buyer to check it. If you need to know what four hundred county courts did with that rule last year, you need the record of what they did, and that is a different artefact with different problems — redundancy, missing labels, uneven coverage — which we publish rather than hide. The teams that ship working China coverage are the ones that stopped asking which source is better and started asking which question they were answering.