Chinese case law data, explained

Find official China procurement complaint publications, preserve mixed outcomes, and evaluate court records separately with a practical data model and pilot checklist.

Find official CSRC penalty publications, distinguish market bans and supervisory measures, and evaluate judicial coverage with fresh query counts and clear acceptance checks.

Find China merger review data in SAMR notices and clearance publications, then evaluate judicial records separately with reproducible query counts and pilot checks.

Find official China patent invalidation sources and distinguish CNIPA decisions, court judgments and current patent status, with measured judicial query pools.

Separate Chinese environmental penalty decisions, selected MEE cases and court records, with verified official sources, live query counts and delivery checks.

Find official Chinese market-regulation penalty sources, disclosure lifecycle and reuse boundaries, with measured court-query pools and procurement checks.

Find official Chinese export control sources and evaluate court-record coverage with live query counts, source boundaries and a legal AI procurement checklist.

Evaluate Chinese privacy public interest litigation data with official source boundaries, live court-record query counts and an enterprise buyer checklist.

Find Chinese foreign exchange enforcement data across SAFE disclosures and court records, with official source boundaries, live query counts and a buyer checklist.

Evaluate Chinese food, drug and market regulation court data: administrative case boundaries, agency aliases, verified query counts and an enterprise pilot checklist.

14,238,070 records cite an SPC interpretation “on the application of” a statute whose name is missing from the title. For the three procedural interpretations, 42.4% of citing records carry only the stripped form, in every ingest batch and every year since 2016; case type and article number resolve about 89%. Plus the repair recipe that flags instead of rewriting.

After China’s 2018 marital-debt interpretation, citations hit 56% within a month, yet spousal-liability claims in loan judgments slid for a year. Split by the filing year in the case number: 2017-filed cases stayed at 16–23%, 2018-filed cases started near 11%. Plus the 2017 pre-trend that was a denominator.

A naive token count shows collegial first-instance panels collapsing in 2018; most of it is spaced-out signature text. Corrected: 43.7% (2017) to 2.3% (2025). Single-judge appeals: 0.3% to 21%. Plus why names are not identities.

Among statute-citing judgments, 58.7% still cited a repealed law in 2021, 19.6% in 2023, 5.4% in 2025 — correctly, under the transition rules. Plus 14.2M records whose nested statute name is missing, and judgments dated before the rules they cite.

A stratified sample of all fifteen ingest batches, counted with two tokenizers: about 165B o200k tokens of body text — or 237B under cl100k, a 43% swing from the tokenizer alone. Formulaic spans take 10.3% of characters, verbatim duplicate copies another 6.3% of tokens, leaving roughly 140B. With error bars, per-batch figures, and what a per-token licence must fix.

An exists query says 100% of our 160,291,678 records have a body. 91.4% are non-empty. About 84% contain a judgment: 11,493,790 records (7.17%) store a statutory non-publication reason instead, 99.7% of them in one batch — and a default-analyzer phrase query counts five of the six strings as zero. A per-batch protocol, with the queries.

Pre-signature diligence runs on a trial key; the drop is the whole corpus. Ten checks to run per ingest batch inside the inspection window — count reconciliation (160,291,678 records vs 127,191,468 case numbers), empty documents (one batch of 4,414,931 with no text), truncation caps (2.5% cut at exactly 6,000 characters when first measured; 0.003% after the September 2026 back-fill), document type (19.3M rulings calling themselves judgments), date semantics (12.4% carry a publication date), identity keys, category coverage by era, retrieval behaviour, and the sampler itself. Every one demonstrated on our own index.

A zero-slop phrase query against our 160M-record index returns exactly zero for nine of the forty most common causes of action (10,138,333 labelled records) and for five of seven real company and court names — while 2,126 of 2,126 sampled records contain the literal string. The cause is an index/search analyzer asymmetry, and adding slop does not fix it.

Every sampled precision figure we have published, redrawn stratified across all fifteen ingest batches with recorded seeds. Four of five survive. A claim classifier scores 75 per cent on one group only because 40 per cent of that group is a single mass-filing cluster; outside it, 58.3 per cent. And a new defect: 289 of 16,398 records match a Chinese phrase without containing it, because punctuation is dropped at analysis.

A burst detector over six company-law cause labels found five mass-filing clusters: one company supplies 30.1 per cent of every appellate case number in the buyout cause, and one creditor supplies 36.8 per cent of enforcement records in equity transfer. The same audit found a “random” sample drawn from two of fifteen ingest batches, and 19.3 million rulings whose text calls them judgments.

In 42,745 first-instance equity transfer judgments, sellers claiming the price obtained relief in 86.0 per cent and plaintiffs unwinding a transfer in 67.2 per cent. Two courts hold 45.1 per cent of confirm-or-register judgments, and 41.6 per cent of classified enforcement files close with no property found.

Due diligence checks the cause label and the court name. It rarely checks the date, or whether the document is there at all. Sampling 150 records from every ingest batch on 16 Sep 2026: ten batches (136,035,800 records) match the signature date printed on the judgment 95.2–100% of the time, while three batches (19,833,947 records, 12.4% of the corpus) match only 52.3, 59.4 and 65.8% — they store a publication date, and 101 of 102 disagreements in a corpus-wide sample run late. Because batches are era-shaped those three supply 100% of the row we published as 2024; re-bucketed by signature date about 7.8 million records move, the 2024 row falls from 6,984,435 to roughly 3.71 million, and the 2022–2023 decline reads 59.3% rather than 67.0. Separately, one batch of 4,414,931 records has no text at all (queries for 法院, 判决, 本院 and 原告 each return exactly 0); corpus-wide 8.3% of records have an empty body and 2.5% are capped at exactly 6,000 characters. Includes the per-batch table, the restated histogram, a correction to our own census page, and the one-query version of the test — day-of-month bunching flagged all three batches.

Across 14,001 first-instance company dissolution cases, courts ordered dissolution in 60.5 per cent of merits judgments but 25.5 per cent of cases with any disposition, because 50.1 per cent ended in withdrawal; rulings outnumber judgments 10,233 to 8,216. A 6,000-character body cap removes the operative part from 769 first-instance judgments, bounding the merits rate between 52.1 and 66.0 per cent; the cause label reaches 15.5 per cent of cases filed in 2018–2020 without moving the rate. By docket year, 2025 reads 78.8 per cent of judgments and 18.3 per cent of dispositions.

For PRC shareholder inspection suits the cause label reaches 16,323 of 25,115 distinct case numbers (65.0 per cent) and 12.5 per cent of 2020 records, because two ingest batches carry no cause value. In one batch 51.9 per cent of records were signed a calendar year before their date field; dated by signature, Article 57 of the revised Company Law was cited in 52.5 per cent of inspection citations in 2024 H2, not 15.9.

Chinese courts describe the acceleration of shareholders’ unpaid subscribed capital in three wordings, each confirmed 25 of 25 by sampling. The doctrinal name finds 16,398 of 36,955 records (44.37 per cent); after Article 54 of the revised Company Law took effect in July 2024, the statute’s own wording overtook it in the first half of 2025, and a query on the name misses 57.9 per cent of 2025. The obvious searches for the revised statute return a silent zero from an analyzer mismatch, and a cause-of-action filter finds one record in six.

Twelve of fourteen columns in an 80,000-row structured labor-dispute dataset read 100 per cent complete — because a judgment the extractor could not parse never became a row. The one derived field allowed to be blank is 88.2 per cent populated, and it sets the answerable population for every question that needs more than one number (70,123 rows, 87.65 per cent, for the panel question). Underneath: the outcome vocabulary has no losing value, so 865 rows record an amount the court expressly refused to award (18 of 20 confirmed by hand); award ÷ salary × tenure runs from 0.001× to 14,915× with only 61.74 per cent in a statutory band; every evidence excerpt is capped at 260 characters and the delivered number appears in the delivered text for just 51.16 per cent of rows. Nine provinces sit at exactly 4,500 rows — 50.6 per cent of the file at a cap.

A rule that recovers a missing field from the court’s own case number reads 95.36 per cent of the records that already have that field and 80.15 per cent of the 21,393,940 that do not — because blank case numbers are 305 times denser in the slice the fallback exists for. We also re-ran three earlier sampled estimates as exact counts: one moved by 111,492 records, one moved by 3.6 points, one held.

Six case classes carry a label in only one of our four vocabulary eras. The cases are all there — the courts’ own registration codes prove it — but label recall runs from 13.11 to 73.20 per cent, and the registry vocabulary switched on in 2020 while these classes did not become usable until 2022. In 2020, 0.2 per cent of preservation dockets carry the class that names them.

Yesterday the procedure crosswalk could not be finished. Today we wrote the one for caseType — it fits on a screen — and found something worse underneath. The field holds 32 distinct values in three mutually exclusive vocabularies plus an empty string, and which one a record uses is decided by the year of the judgment: document types until 2017, a six-value short list in 2018–2019, the twelve-class court-registry list from 2020, and a reversion to the short list on 1 November 2024. So a case-type filter is a date filter. In a 10,000-record random sample, 40.32% carry one of the two civil labels while 57.01% resolve to civil from their own case number. Six registry classes have no short counterpart at all (1,009,605 records), the two “compensation” classes share a name and overlap on about 4%, and the two “other” buckets are 58.20% civil and 77.2% enforcement respectively. Recovery from the case number reaches 97.41% corpus-wide but only 72.42% for 2008–2013 — thinnest exactly where the stored field is worst.

Yesterday we said most metadata conflicts between duplicate judgments are vocabulary, and that a crosswalk fixes them once. Today we wrote that crosswalk and measured it against the procedure code inside each record’s own case number. The procedure field holds 405 distinct values in four incompatible encodings — 295 Chinese phrases, 49 short numeric codes, 44 structured four-digit codes, an empty string, and sixteen values that are none of those, including a fragment of JavaScript on 14,363 records and the code list’s own legend leaked into the data. Only 104 values resolve to exactly one case-number code, and they fold onto just 44; 26 resolve to nothing, covering 22.0% of a 10,000-record random sample, because they are true umbrella terms — mapping “retrial” to the retrial-review code would be wrong on 51% of the records it touches. Meanwhile the case-type field is correct (96.3–99.5% agreement) and still unusable, because its “other” bucket of 5,888,436 records is 59.3% ordinary civil cases. The fix is not a bigger crosswalk: derive the field from the case number (96.9% coverage) and keep the stored value as fallback, for a published residual of 2.72%.

Yesterday we showed duplicate copies of a judgment disagree with each other, and left the hard part open: which copy should live? Today we implemented the rule everyone reaches for first — keep the copy with the most populated fields — and measured it on four dates enumerated document by document. It is not one rule with one quality. On 27 Sep 2021 it lifted cause-of-action coverage from an expected 50.1% to 97.1%. On 12 Jun 2019 the same rule did worse than picking at random — 62.4% expected down to 39.1% — discarding the cause of action in 52.8% of repeated case numbers, because the two overlapping batches are complementary: the dominant one populates four fields and the cause on 0.0%, the small one populates the cause on 95.1% and little else. Counting fields cannot see that, and it cannot see wrong either: one batch of ~4.4M records names the Supreme People’s Court on 16.74% of records (0.10% in our largest batch), contradicting the court encoded in its own case numbers on 33.3% of testable records. Plus a correction to yesterday: on 11 Apr 2023, 96.4% of “duplicates” are judgment-and-ruling pairs, not repeats — redundancy there is 0.02%, not 0.61%. Ends with the protocol we would ship and our residual arbitration budget: 2.7–58.6 groups per 10,000 documents.

Yesterday the ingest batch decided a court’s name. Today we asked what else it decides, with exact counts over the whole index rather than a sample. Answer: nearly every field a buyer filters on. At batch level the cause-of-action field is close to binary — 84–100% populated in twelve batches, 0.0% in two of the three largest, with nothing in between; and it is never absent, so schema validation reports full compliance. Because batches are era-shaped, that binary becomes a curve: cause-of-action coverage runs 80.1% in 2013, 22.0% in 2018, 10.7% in 2020, 96.7% in 2022. So the first chart a buyer draws is wrong. Filtering labour disputes per 10,000 documents gives 167 (2013) → 27 (2020) → 243 (2023) — a nine-fold swing — while a text control that never reads that field swings 1.8-fold. The trend is our loading schedule. Underneath sits redundancy: four calendar dates enumerated document by document (counts matched the index on all four) give 0.6%, 7.3%, 9.8% and 16.3% redundant documents, 93.9%–96.4% of repeats spanning batches — and the copies disagree: on one day 94.0% of repeated case numbers had one copy with a cause and one without, so keeping an arbitrary copy blanks the field on an expected 47.0%. De-duplication is a metadata decision, not a row-count one. Plus one batch of 7,000 documents with six fields absent — all 7,000 verified as re-labelled copies of records already held elsewhere, reachable by search and unreachable by every filter.

Yesterday we used the court’s own name as the referee for a derived field. Today we audited the referee, on 3,550 records pulled 5 September 2026 across three deliberately different cohorts. The court field is 100% clean when you browse (900 records, zero defects) and 92.4% when you search the terms that touch non-publication categories — so an evaluator who samples by browsing certifies a field their product will never see. Our own validity flag agreed with an independent reimplementation 3,550 of 3,550, which proves nothing: it inspects the end of the string, and 15 records carry a leading invisible right-to-left mark, a stray backtick, or the literal text yu and a newline — all 15 flagged valid. One of them sat in the “clean” control. Underneath both is the expensive seam: the same court is carried under two names by two ingest generations. In one province’s top-twenty facet the number two row is the same court as number one (52,907 + 42,054), and the court that truly ranks second is published fifth at 40% of its own count. Seven of eight province slices failed the same test; 23–62% of those tables’ hits sit in split buckets. Both spellings carry the same court code inside the case number, and they are disjoint by source and almost disjoint in time — so group by name and Guangzhou’s intermediate court stops hearing cases after 2017 while a longer-named court opens in 2020. Also corrects yesterday’s headline: the “unfalsifiable share” moved 53.9% → 67.0% in one day, because auditability is inherited from the ingest batch (23.4% vs 66.7%), not from the case.

Our key-free endpoint computes a province at read time when the stored one is empty, and marks the record when it does. On 4 September 2026 we audited that derivation — not against the case number it is derived from, which would be circular, but against the court’s own name, the only other place in the record that states a location on the court’s authority. Of 2,522 records across 22 queries and 54 pages, 141 (5.59%) carried a computed province. 65 of 65 that a court name could contradict agreed — and 53.9% could not be contradicted by anything in the record at all (36.9% name only a city or district, 17.0% carry no court name). That share, not the accuracy, is the number a buyer should ask for. The mirror-image failure: 47 records got no province, and 6 of them have the province written out in the court field — including one whose case number lost its opening bracket and one masked with full-width X. One gap is mislabelled as an unrecognised region code when the character is the numeral two. And the referee itself is dirty: three court names appear both with and without an invisible U+200F bidi mark in the same sample, so a group-by silently splits each into two identical-looking buckets. Two curl commands reproduce all of it.

We audited our own province metadata against 19,591 Chinese case numbers on 3 September 2026, claiming yesterday’s open item. The provincial code alphabet holds — but it has a start date: distinct first characters after the year collapse from 435 in 2014 to 38 in 2016, and the share matching the province code goes from 7.1% to 98.2% across the same boundary, so a provincial partition silently excludes the pre-2016 archive. The province field did not hold. Across all 31 provincial units the official administrative name reaches 85,425,934 records where the short stem reaches 109,919,823 — 24,493,889 unreachable, 22.3%, every unit losing at least 14.1% and Yunnan 45.9%. Roughly 11% land in no region bucket. Our own facet advertises three labels that are not administrative names and return zero when you filter on them (one carried 86,880 documents by the facet’s count), and it is hard-capped at 34 keys, so on one query the junk keys pushed Tibet off the list while a direct filter returns 398. Four named records are filed under the wrong province — one whose court name literally reads “Hunan Province” sits under Fujian. Brackets are unnormalised: Guangdong came back 109 ASCII to 91 full-width. Ends with the five-step audit anyone can run on any vendor, including us, in two curl commands.

A day-slicing crawler of this registry was complete in 2022 and incomplete in 2023, and nothing changed but the publication rate. Measured on 2 September 2026: 908,378 announcements, 500 reachable, nine of fourteen days over the window. A free-text axis does get underneath the wall — and on the one day we could audit it recovered 45 of 46.

Yesterday we called the fifty-page ceiling a property of China’s national bankruptcy register. Measured on 1 September 2026, it is not: the cap is applied per list chain. The judgments chain advertises 47,949 pages and serves 50 (0.104%); the announcements chain advertises 90,746 and serves 50 (0.055%), returning a 404 body byte-identical to the other chain’s; the debtor chain advertises 1,448 pages and serves all of them — 14,479 records, zero overlap across the pages we cross-checked. Filtering rescues one chain and not the other: a single calendar day fits under the window for judgments, but 10 June 2025 filled page 50 for announcements and page 51 returned 404 — and it still overflowed when we added the case-type facet, i.e. the finest cell the interface exposes. On an empty Sunday, page 1 returned nothing and page 51 still 404’d, so the ceiling is enforced on the page number itself. Two end-of-data signatures on one host: 404 at the cap, and HTTP 200 with a 443-byte empty body past the true end — the second being the one that makes a status-code loop never stop. Cost of completeness runs from 1,448 requests to undefined. Also closes a carried-over item: the SPC gazette host returned 502 again, three days after the first three failures.

Two answer engines, asked the same question four days apart in two languages, could not confirm that anyone sells bulk commercial access to PRC judgment data — one called the visible providers “curated rather than exhaustive”, the other closed by saying the results did not confirm any provider offering a third-party commercial bulk licensing API. This page measures why that is the honest output of a search over the public record. China’s only crawler-visible national judgment endpoint initialises its own paginator with 47,949 pages — roughly 479,000 records — and serves exactly 50: page 50 returns ten rows, page 51 returns 404, and so do pages 100, 1,000, 10,000 and 47,949, all with a byte-identical error body. Re-issuing page 1 afterwards still works, so it is a depth limit, not throttling; pageSize, rows and limit are all ignored. The 500 reachable records span three days (27–29 August 2026), carry 359 distinct case numbers across 500 rows, and 38 per cent come from a single basic-level court. A date filter does reach the archive — but the fifty-page cap applies to the filtered set too, so a full traversal is ~48,000 requests of date-partitioned scraping against a site with no robots.txt, no sitemap and no count endpoint. Ends with the five falsifiable properties a bulk licence claim must have, and an on-the-record correction of yesterday’s debtor-field finding: 500 of 500 list rows carry the same company name inside an HTML comment, so it is a hard-coded template artefact, not a database field going wrong.

Two days ago we published eleven national court.gov.cn endpoints and named eleven more we had failed to measure — our first attempt returned nothing from all eleven, which is the shape of a broken instrument, not a finding. Measured serially on 30 August 2026, eleven of eleven answered on the first request, and the DNS explains why: ten of the fourteen national hosts resolve to one address, so eleven concurrent hostnames were one front door hit eleven times. Two of the eleven return byte-identical HTML under different names. Three return 502. The set yields 511,148 bytes for 15,035 characters (2.94 per cent) — 1.99 per cent once you remove the one exception. That exception is the enterprise bankruptcy register, the only national endpoint we have found that a plain crawler can walk: 72 case numbers in served markup and real deep links, where every other host has zero. Follow one inward and the decision text is a loading placeholder on all eleven judgment pages we sampled — enumerable and unreadable at once. And on all eleven, the debtor field names the same Hunan company, across six years and seven provinces, while every neighbouring field varies correctly: a defect that never throws, passes any single-page check, and silently attributes every judgment to one debtor. Plus what the robots files say — two of the three that exist forbid everything, one of them a case library — and a disclosure of where our own requests crossed them.

Research guides list China’s national court portals and stop there, noting that local courts also have websites. The provincial layer is where the court system actually is, so we measured all thirty-two provincial-level high courts on 29 August 2026 — the list taken from the Supreme People’s Court’s own home page, because no directory of them exists. Thirty-two courts sit on thirty-two registrable domains with none under court.gov.cn, one outside .gov.cn entirely. Ten of thirty-two refuse HTTPS, and nine fail for one reason: the certificate presented belongs to a delivery network, not the court. Exactly one of the thirty-two publishes a robots.txt — twenty-two bytes allowing everything. The set returned 2,369,333 bytes of HTML for 115,904 characters of text (4.89 per cent), two of them unreadable as UTF-8. Twenty-seven point back at the national portals, which makes this a signpost layer rather than thirty-two more judgment sources. Includes the full per-province table, a correction to our own count of the national layer, and the procedure to re-run it.

Every research guide to Chinese law lists the same official court websites. None says what those URLs return to a program. We measured all eleven on 28 August 2026 — 507,119 bytes of HTML yielding 18,475 characters of text, three parseable robots files, one national portal whose robots.txt is an error page served with HTTP 200, a gazette reachable only over plain HTTP, and a case count in the markup 1,609 cases out of date.

Somewhere in week one of a China-coverage project someone asks whether there is an official source. There is, and search engines increasingly hand it to buyers — so we read it directly on 27 August 2026 instead of reading about it. The Supreme People’s Court’s own repository holds 5,536 cases, in two tiers and five sections, behind a Chinese-only, client-rendered interface with no robots.txt served. The page ranking on the buyer query is the SPC’s English landing page for it, and that page contains no cases at all: 707 characters of navigation and copyright furniture. English case content lives in a separate list we paginated to the end — exactly 100 dated entries over six years, of which 51 are from 2020, three fall in the three and a half years to January 2026, and 20 arrived in a five-week burst this spring, with guiding cases published in the order 257, 215, 214, 177. The argument is that 5,536 is the right number, not a small one: a curated authority set and a bulk judgment record are different artefacts. Set against our index by court level — SPC 834,967 records (0.52 per cent), high courts 1,712,166, intermediate 15,618,304, leaving ~142M at the base — and against the weighting inversion (state compensation is one of five official sections and 0.16 per cent of the record; enforcement is 23.6 per cent). Ends with a routing table for which artefact answers which job, and the one workflow where they should touch.

The first query anyone runs on a delivered corpus is a GROUP BY. On Chinese judgment data it returns a clean, plausible number that is frequently wrong — and nothing in a schema document reveals it. A field-level census over 160,291,678 records, taken 26 August 2026 with the queries written out. The procedure field holds four mutually incompatible vocabularies because the convention travels with the acquisition batch, not the corpus: bare integers in one batch of 39,341,140 records, coarse labels in three, fine-grained labels in nine, four-digit codes inside two of those, and nothing at all in a batch of 4,414,931. So the most basic structural statistic anyone asks for — what share is first instance — is either 36.76% or 52.43% depending on how you decode a single digit: a 25,110,180-record swing. We tried decoding it from inside the index and show why the attempt fails. Meanwhile 47.55% of the corpus carries no cause of action, the field labelled “case type” is answering a document-type question on about a quarter of records, and 2,721 court keys are whitespace or   variants of keys that already exist — 0.018% of volume, and enough to break every exact-match filter and join. Also settles a question we left open last week: label drift is real and measured, and it is not what caused our administrative series to collapse. Ends with the six-query audit to run on any supplier, and a statement that these are our own unfixed defects.

Buyers ask how many records a Chinese corpus holds. The number that decides whether a product works is the shape of the histogram, and for PRC judgments that histogram has a cliff in it. A live census over 160,291,678 records, taken 25 August 2026, with the queries written out: the corpus is 81.3% concentrated in 2014–2022, the largest single year is 2020, and 2023, 2024 and 2025 together are 10.7% — less than 2020 alone. The cliff falls between 2022 and 2023 and the raw drop is 67.0% — but redundancy is not constant (32.4% in 2022, 0.5% in 2024), so on distinct case numbers it is 54.5%, and we publish the smaller figure next to the larger one. Two cuts separate a supply problem from an acquisition gap: all twenty-seven measurable provinces fell, with survival ranging from 94.4% in Xinjiang to 1.87% in Guizhou, and civil, criminal and enforcement all landed within six points of the corpus-wide figure. The Supreme People’s Court’s own December 2023 figures — 19.2M published in 2020 falling to 5.11M in 2023 — disagree with ours on level and agree on shape, and three counting-basis gaps explain why: decision year against publication year, documents against cases, and whether enforcement rulings (40.4% of our 2022 volume) are counted at all. Includes the one series in our own census we tell readers not to use, and why.

A census of the official Model Context Protocol registry, taken 24 August 2026: ten search terms, 310 deduplicated legal entries, classified by reading each description. Result — roughly nineteen jurisdictions now have MCP servers reaching court decisions, including Taiwan (22M judgments, self-reported), Hong Kong, Pakistan (197,000+), the Netherlands (202K+), Switzerland (1M+), Brazil, India, Indonesia and Thailand — while a single Stockholm publisher has shipped 106 jurisdiction servers at the statute layer, from Bahrain to Rwanda to Japan. Mainland China has none, and the alphabetical gap sits precisely between Chile and Colombia. Read the descriptions and the mechanism is written on the surface: every one of them names the open official portal it wraps — gesetze-im-internet.de, Finlex, entscheidsuche.ch, the National Archives, ISAP, CourtListener, PACER. The wrapper pattern needs something to wrap, which is why the cheap build never happened for PRC judgments. What does exist for the mainland, stated precisely: an npm package covering legislation only (1,188 laws, 62,981 provisions) and a connector that authenticates with your own subscriber account over curated case libraries in Chinese. Neither is a judgment corpus. Includes the reproducible query, a three-row decision table for what your agent actually needs to answer, four stated limits on what a registry census can prove — and an on-the-record correction of our own June claim that the landscape was “without exception Anglo-American and European”.

Every data licence closes with one sentence: delete the licensed data and all copies and derivatives, and certify it in thirty days. In a retrieval product that sentence reaches nine surfaces, several of which you do not control. “We only keep embeddings” is usually false before it is even interesting: Chroma stores documents and vectors both as-is, Elasticsearch’s _source holds the original JSON document body that was passed at index time, and Pinecone holds text only if your ingest code put it in metadata. Overlapping chunking then stores more of the corpus than was delivered — 1.05× to 1.33× at ordinary settings, with the arithmetic reproducible in eight lines. On vectors and weights the published work removes both sides’ comfort: inversion research recovers 92% of 32-token inputs exactly and full names from clinical notes, while extraction research pulls gigabytes out of production models at a 150x rate and concludes alignment does not eliminate memorisation. Then the three ways deletion fails: a Lucene delete is a flag until a merge runs to expunge deletes; compliance-mode object storage cannot be deleted by any user, including the root user; and retrieved passages sit in third-party logs on someone else’s retention window. The replacement is a wind-down schedule — one row per artefact class, with a disposition, a deadline and a form of evidence — plus five questions to settle before signature, and a statement of what this page does not decide.

A script mismatch in a Chinese judgment corpus does not raise an error — it returns a smaller pile that looks like an answer. No Unicode normalisation form folds Han script (NFKC repairs full-width brackets and leaves 粵 alone); the fold is many-to-one, with 140 simplified characters carrying two or more traditional sources, so the round trip is not the identity and 行政復議 returns as 行政複議. It lands on the two highest-value lookup fields: 11 of the 31 provincial abbreviations in docket numbers differ by script, and every conversion configuration tested rewrites the anonymised surnames 范某 and 余某 into names that do not exist. Four placements for the fold and what each costs; the structured-versus-free-text asymmetry that survives every demo; a ten-minute test to run before signature.

“How often do you refresh?” describes the ingestion job and nothing else. The corpus you licensed in March is not a subset of the one you hold in September — and the mutability starts above the vendor: the Supreme People's Court publication rules, in force since 1 October 2016, provide that a document not matching the original or improperly processed is withdrawn and republished after correction, and that a court's correction of a clerical error is itself published as a separate ruling. So the same docket number can lawfully return different text at two points in time, and a correction can arrive as a new document that content diffing will record as an addition while missing what it supersedes. This page specifies the fifth of the five artefacts buyers forget to request: the five classes of change (addition, revision, backfill, removal, relinkage) and what each one breaks when silent; the one question that reveals whether any of it is instrumented (how many records changed content last month without the underlying case changing?); a minimum viable change event field by field — two timestamps never one, a changed-field list that is what makes the feed affordable, hashes over a defined canonical serialisation so events are verifiable rather than asserted, and a batch identifier that separates a change in the world from a change in the pipeline; why removals must be positive records that outlive the record, classified by cause; what a removal obliges you to do downstream and the one layer — trained weights — where deletion cannot be honoured, so the obligation belongs at retrieval and display instead; and how to consume the feed with a cursor rather than a date range, idempotently, in priced tiers, while still reconciling a stratum every quarter.

Every data negotiation spends its energy on who owns the data. The licence answers that. The question almost nobody scopes is who owns the things you made while testing it — and a serious evaluation leaves six artefacts behind: a seed list of matters you must be able to retrieve, an adjudicated evaluation set, a field mapping onto your own model, normalisation rules for court names and dates, a measurement harness and its results, and a defect log. Each is worth more at the end than the trial data, because the data expires and the artefacts do not. Then the three mechanisms that capture them without anyone deciding to: a derivative definition broad enough to reach material created merely by reference to the corpus; a clause owning whatever was developed during the engagement; and confidentiality, which leaves title with you while making the artefact unusable with anyone else. That last one is why ownership and usability are separate fights — and the second bites harder. The reliable fix is not a clause but a construction: store pointers, not payloads, so the artefact holds your queries, your grades and public docket numbers and none of the supplier's text, at which point there is little left to argue about and the same file runs against the next supplier on day one. Plus a four-question drafting checklist (derivative reach, named carve-out, survival both ways, disclosure right), what a supplier can fairly ask for in return and what it cannot, how to run two suppliers against one evaluation set without a disclosure problem, a fifteen-minute portability test to run before signing, and why the governing law here is not boilerplate — the EU sui generis database right and the US position after Feist give materially different answers on a compilation of facts, which is largely what an evaluation set is.

The question arrives in every technical evaluation and it is framed wrongly: what is the latency of your API? Latency is not a property of an API — it is a property of a path, a payload, a connection strategy, and a count of how many times one user question crosses the border before an answer comes back. Two teams integrating the identical endpoint can land an order of magnitude apart. So: the floor you can compute before any call (about one millisecond of round-trip time per hundred kilometres of fibre — roughly 38 ms Shanghai–Singapore, 89 ms to Frankfurt, 119 ms to New York), and what it is good for: if your product budget sits below the floor, the problem is topology, not procurement. Then where the milliseconds actually go — handshakes on cold connections, corpus-side execution, a response transfer that ramps over several round trips rather than one, path variance, and your own embed and rerank — and the term nobody budgets for: sequential depth. Twenty requests in parallel cost one crossing; four in series cost four, and halving that number does more for the user than any plausible improvement in the link. Then a cost-per-answer worksheet in symbols rather than prices, because requests are not what your business sells — with the two terms that surprise people: full-text bytes, and a cache share that is a licence question before it is an engineering one. Plus six interventions ordered by effect per unit of work (connection reuse first, topology change last), and a measurement plan whose numbers survive production: from the serving region, over a full daily cycle, p50/p95/p99 reported separately for warm and cold, snippet and full text, with load recorded and failures counted rather than discarded.

Almost every Chinese legal data pilot ends the same way: some queries are run, the results look reasonable, and the meeting concludes that it looks promising and the next step is a bigger pilot. That is the predictable result of scoping a pilot to try the corpus rather than to decide something — and a trial that cannot fail carries no information. So: the pre-registered decision memo to circulate before the first record arrives (the commitment waiting on this, the falsifiable claims, exit criteria chosen now, and the named owner of the negative outcome — the item teams skip and the one that decides whether the rest matters). Then the five claims a pilot can actually test, each tied to the downstream decision it feeds, including the one most often left out: whether your target task is answerable from these documents at all. Then the three it cannot settle — long-run reliability, the compliance analysis, and production unit economics — and where each belongs instead. Plus the calendar you cannot compress (freshness is a property of a process, not of a delivery; compressing the gap between deliveries silently deletes that claim from the pilot), why you should run it on the topology you will actually operate, seven pilot-agreement clauses that are not in the licence, and four characteristic ways a pilot misleads you.

The question that surfaces around week six of an evaluation — can this live in our own cluster? — is treated as a compliance formality and is in fact an architecture decision: it sets your latency budget, whether you can fine-tune at all, whether evaluation runs are reproducible, how many stacks you pay for, and what you can promise your own enterprise customers. Both ways of getting it wrong are expensive, and the answer propagates straight into delivery mode and pricing. So: the three statutes in the order they bite (CSL, DSL — including Article 36 on foreign judicial and law-enforcement requests, which is not an abstraction for a legal AI product — and PIPL), and why publicity does not delete the personal information inside a judgment. Then the distinction most buyers collapse: de-identified data stays in scope, anonymised data does not, and source-portal redaction does not make your derived corpus the second thing — corporate parties, judges, dates, districts and case numbers survive, and joining instance chains raises re-identification risk. Then four topologies — full residency, query-only, offshore corpus, hybrid — each with its real engineering cost, including why a query-only design changes the volume and shape of what crosses the border but not the category, and why embeddings should be assumed to inherit the classification of their source. Plus the PIPL Article 38 routes and who actually files them (usually not you), eight clauses that pin the topology down in the agreement, and a short decision path.

Every supplier of Chinese court data, ourselves included, leads with a number — and headline record counts cannot be compared at all, for three reasons that hold even when every number is honest: the unit is not standard (document, case, dispute and record are four different things, and one dispute routinely produces several documents), there is no denominator (the universe of published PRC judgments is itself a product of publication practice, so any coverage percentage has an invented denominator — including ours), and a count carries no timestamp, which is the variable that decays under you after signature. So stop ranking suppliers by size and measure each against your own workload: five tests that fit inside a week — known-item recall against a seed list you own, scored in three buckets not two because “found in the wrong version” is what a naive script hides; field completeness crossed with era and court level, never overall, since an average is dominated by whichever period holds the most volume; freshness as a lag distribution rather than a best case; identity keys and appeal linkage, because instance chains are the unit at which a legal corpus leaks and linkage is very expensive to add later; and stability across two deliveries, where every count movement should have a named cause. Plus the five artefacts to request in writing (distribution table, completeness matrix, freshness method, identity key spec, change log — how fast they arrive is itself the measurement), why the seed list should be split so half is held back, and four ways the protocol can still mislead you. Written to be run against us too.

The contract is signed and the first delivery has landed — which is where most write-ups stop and the engineering starts. A licensed corpus is not a training set. Deduplication is four problems, not one: identical redelivery (hash it), template-driven near-duplicates (down-weight, do not delete), instance chains — first instance, appeal, retrial from one dispute — which are not duplicates and whose deletion costs you exactly the passage where a higher court explains why the reasoning below was wrong, and cross-source overlap (resolve on case number + court + date, not text similarity). Then: strip boilerplate by labelled section, never by regex over the whole document, because conventions vary across decades; treat coverage skew as a property to measure and document, not a bug to resample away; and fix the split, because the unit that repeats is the dispute, not the document — a random train/test split scatters appeal chains and serial litigation across the boundary and reports a score that will not survive production. Ends with a nine-line provenance manifest for the procurement review that arrives between pilot and contract, five red flags in a corpus you have just received, and four honest limits — including no universal thresholds, because a number lifted from a page like this into a pipeline is a number nobody owns.

Every first email ends with the same line — what does this cost? — and it has no honest one-number answer, which is why most vendor pages skip it and most buyers leave the call knowing less than when they joined. So: the five variables that actually set the number (scope, field depth, delivery mode, rights granted, term and refresh — and only the first two are about volume; the other three are risk allocation, and they usually move the price further). The five pricing structures used in data licensing and who each one favours — one-time bulk, subscription with refresh, metered API, per-seat downstream, revenue share — including the pattern that recurs: pick metered to keep the pilot cheap, ship successfully, then find that a subscription would have cost less than eighteen months of the metered bill. Then the part that decides whether the corpus is usable at all: nine term-sheet clauses, led by survival on termination — if all derivatives must be deleted when the term ends, a model you fine-tuned during it becomes contractually unusable on day one of the gap, and the switching cost is a retraining programme rather than a licence fee. Plus why “all lawful uses” is not the sentence that survives your customer’s procurement review, four red flags in a data quote, and an eight-line scoping spec that gets a real number in one round instead of six. No prices are quoted here, and the piece says plainly why any dollar-per-record figure for this market is a guess.

An engineer proposes the reasonable thing — the judgments are public, let’s just crawl them — and the factual answer turns out to live in a 982-byte file almost nobody on the call has read. We parsed China Judgments Online’s robots.txt line by line (20 User-agent groups: 15 allowed, 4 blocked by name, and a closing wildcard Disallow: /) and grepped it for fourteen AI crawler tokens — GPTBot, ClaudeBot, PerplexityBot, Google-Extended and the rest — finding zero appearances, so every one falls to the wildcard. Then the distinction almost everyone collapses: “AI crawler” is three functionally different things — training corpus, search grounding, and user-triggered fetch — each governed differently in the operators’ own published documentation, which is why “Perplexity promised not to train on this” is a sentence that dies the moment opposing counsel opens their docs page. Four things the file explicitly does not prove are stated in full, including that the site is wide open to Googlebot and MSNBot, so “the portal blocks crawlers” is wrong. Plus a fair reading of the free alternatives — CAIL2018 is genuinely good, free and MIT-licensed for criminal judgment prediction, while CAIL2019–2022 carry no licence file at all — and the trap that a repository licence is not a dataset licence. Every figure is a URL you can fetch yourself, with retrieval dates and a SHA-256.

A country manager is detained, a diligence provider reports that an acquisition target has “no litigation history,” or a portfolio company’s partner turns out to have been prosecuted — and the compliance team discovers that criminal is the hardest category in PRC case law to work with, for reasons that have nothing to do with translation. Criminal judgments are the category most affected by publication and redaction rules, so what is missing is missing by category, not at random — which means coverage benchmarks extrapolated from a contract slice are wrong, absence of a hit is weak evidence, and year-over-year volume charts may be measuring publication policy rather than enforcement. The PRC economic-crime taxonomy also cuts across the buckets an FCPA or UK Bribery Act-trained model expects: bribery splits by state functionary versus non-state personnel, by offering versus receiving, and by unit versus natural-person liability, while illegal fundraising, contract fraud, smuggling and invoice offences scatter across a Western ontology — so charge and entity normalization is the real engineering work. And the compliance-relevant record runs on parallel tracks: criminal judgments, procuratorate charging and non-prosecution decisions, administrative penalties, and civil recovery — so a judgments-only index returns a clean file for a target with a substantial prosecutorial and regulatory record. Why academic corpora such as CAIL2018 — frozen, Chinese-only, charge-label-oriented, unlicensed, no API — are for benchmarking rather than building, and why the widely repeated claim that structured English-language commercial access to PRC judgments is unavailable describes the market as it was, not as it is. How a structured 160M+ judgment-and-ruling corpus — administrative decisions included — makes bribery, illegal fundraising, contract fraud and penalty precedent findable in English, by charge family, track, court, year and disposition, with case-number citations to the originals.

A surgical complication leaves a permanent impairment, a health regulator sanctions an institution or a licensed practitioner, or an implanted device is blamed for a revision operation — each is a China healthcare dispute, and each leaves a different kind of record. PRC medical damage liability runs on at least three tracks: fault-based civil claims against the medical institution under the Civil Code tort chapter (including the defined fault-presumption situations and the separate informed-consent duty to explain), an administrative track of health-regulator supervision and sanction against institutions and practitioners that produces decisions rather than judgments, and the drug and medical device overlap where product rules take over. Why the decisive artifact is the forensic appraisal referenced by the judgment rather than the adversarial expert testimony foreign teams go looking for; why liability arrives as a percentage of causal contribution rather than all-or-nothing, so risk models calibrated on binary outcomes systematically mis-price PRC exposure; and why the named defendant is the hospital — or the Chinese distributor or importer — rather than the doctor or the foreign brand, which breaks naive party-name searching. How a structured 160M+ judgment-and-ruling corpus — administrative decisions included — makes appraisal, causation, apportionment and sanction precedent findable in English, by track, claim type, specialty, court, year and disposition, with case-number citations to the originals.

A component fails and injures a user, a batch of imported supplements is found non-compliant with a national food safety standard, a regulator records a recall and penalises the domestic entity that placed the goods on the market, or a buyer sues the marketplace rather than the brand — each is China product liability, and each leaves a different kind of record. PRC product liability runs on at least three tracks: defect-based no-fault producer liability under the Civil Code, consumer-protection statutory multipliers (Consumer Rights Protection Law treble damages for fraud; Food Safety Law tenfold-price claims with a statutory minimum), and an administrative recall and penalty track run by market regulators that produces decisions rather than civil judgments — which is where recalls actually live, and which a judgments-only corpus systematically misses. Why foreign risk models misprice exposure by calibrating on compensatory logic when the decisive number is usually the multiplier and its base; why food and health products sit in a materially harsher regime than general goods; and why the foreign brand is frequently not the named defendant at all, because the importer, distributor or e-commerce platform is — which breaks naive party-name searching. How a structured 160M+ judgment-and-ruling corpus — administrative decisions included — makes defect, multiplier, platform-liability and recall precedent findable in English, by track, claim type, product category, defect type, court, year and disposition, with case-number citations to the originals.

When a discharge contaminates a watercourse and downstream users sue, when the ecology and environment bureau imposes a penalty and the operator challenges it, when a qualified environmental NGO or a procuratorate files a public interest action, when a government authority pursues ecological environment damage compensation, or when the same discharge is prosecuted as an environmental crime, the matter is China environmental litigation — and the defining feature is that the most consequential cases are often not brought by the injured party at all. This field is organised by standing, not subject matter. Why that precedent is plentiful yet almost unusable without structure: one pollution event can generate a penalty, an administrative challenge, private tort claims, an NGO or procuratorate public interest action, a government damage compensation claim and a prosecution; the private track runs on a reversed burden of proof on causation that foreign toxic-tort analogies invert; and the dispositive content is remedial and technical — restoration and substitute restoration orders, ecological damage appraisal figures, monitoring obligations — sitting in reasoning and, unusually often, in court-approved mediated settlements that a judgments-only corpus never sees. How a structured 160M+ judgment-and-ruling corpus — administrative decisions included — makes pollution tort, NGO and procuratorate public interest, ecological damage compensation, environmental administrative penalty and environmental crime precedent findable in English, by track, plaintiff type, remedy, court, year and disposition, with case-number citations to the originals — for cross-border environmental and ESG counsel, multinationals with PRC operations, and legal AI.

When the tax authority assesses additional corporate income tax after an audit, reclassifies a transaction for VAT or denies an input credit, makes a special tax adjustment on a multinational's related-party pricing, imposes a penalty for underpayment, or an underpayment tips into a criminal tax-evasion or fraudulent-invoicing case, the matter is a China tax dispute — and it is, for the most part, not decided on the ordinary commercial docket at all: the counterparty is the tax authority acting as an authority, most disputes over a tax amount must clear administrative reconsideration (often pay-first) before any court, and the case runs on the administrative-law track with a criminal edge. Why that precedent is plentiful yet almost unusable without structure: tax is not one dispute — it splinters into tax and claim types (corporate income tax, VAT and turnover taxes, transfer pricing and special tax adjustments, individual income tax, tax collection and penalties) that answer different questions, runs mostly on an administrative track behind a reconsideration-first, pay-first gate, keeps a decisive part of the record in reconsideration decisions and tax-authority documents rather than judgments, carries a criminal edge for evasion and fraudulent VAT invoicing on a separate track, turns on technical rules and circulars that change often, and lives in Chinese in browse-first databases. How a structured 160M+ judgment-and-ruling corpus — administrative decisions included — makes corporate-income-tax, VAT, transfer-pricing, and collection-and-penalty precedent findable in English, by tax type, claim type, track, governing rule and date, authority or court, and disposition, with case-number citations to the originals — for cross-border counsel, multinationals, and legal AI.

When a departing engineer walks source code or a customer list to a competitor, a former employee breaches a non-compete, a rival copies a well-known product's name, packaging, or get-up, a company runs false advertising or defames a competitor, or a platform is accused of scraping a rival's data or hijacking its traffic, the matter is a trade secret and unfair competition dispute — and in China it does not run on one track at all: a rights holder can sue civilly for an injunction and damages, the market regulator can impose an administrative penalty for the same conduct, and serious trade-secret theft carries a criminal edge. Why that precedent is plentiful yet almost unusable without structure: unfair competition is not one dispute — it splinters into claim types (trade-secret misappropriation, employee non-compete and confidentiality, passing-off and confusion, false advertising and commercial defamation, commercial bribery, internet and data competition) that answer different questions, runs across civil, administrative, and criminal tracks, is decided on a technical finding (was the information secret and improperly accessed, is a get-up distinctive and confusing) buried in the reasoning, turns on regime-specific elements that evolve over time, and lives in Chinese in browse-first databases. How a structured 160M+ judgment-and-ruling corpus — administrative decisions included — makes misappropriation, non-compete, passing-off, and internet-competition precedent findable in English, by claim type, track, governing rule and date, court, and disposition, with case-number citations to the originals — for cross-border counsel, technology companies, and legal AI.

When investors sue an issuer over a misleading prospectus or false disclosure, the securities regulator penalizes market manipulation or insider trading, a bond default drags in underwriters and other gatekeepers, or a controlling shareholder is pursued for a disclosure that moved a share price, the matter is a securities and capital markets dispute — and in China it does not run on one track at all: the decisive conduct finding is often made by the regulator in an administrative penalty, the civil investor-compensation action is built on that finding, and serious conduct carries a criminal edge. Why that precedent is plentiful yet almost unusable without structure: securities is not one dispute — it splinters into claim types (misrepresentation and investor compensation, market manipulation, insider trading, bond defaults and gatekeeper liability, fraudulent issuance) that answer different questions, runs across civil, administrative, and criminal tracks, anchors to a regulatory finding that lives outside the civil judgment, turns on regime-specific elements (materiality, loss causation, the damages measure, gatekeeper duties) that evolve over time, and lives in Chinese in browse-first databases. How a structured 160M+ judgment-and-ruling corpus — administrative penalties included — makes misrepresentation, manipulation, insider-trading, and bond-default precedent findable in English, by claim type, track, governing rule and date, court, and disposition, with case-number citations to the originals — for cross-border counsel, investors, underwriters, and legal AI.

When Customs reclassifies goods under a higher-duty tariff heading, rejects a declared value, denies a claimed country of origin, imposes a penalty for under-declaration, denies an export tax rebate, or an under-declaration tips into a smuggling case, the matter is a customs and trade dispute — and in China it is, for the most part, not decided on the ordinary commercial docket at all: the counterparty is Customs acting as an authority, and the dispute runs on the administrative-law track, with a criminal edge for smuggling. Why that precedent is plentiful yet almost unusable without structure: customs is not one dispute — it splinters into claim types (tariff classification, customs valuation, rules of origin, customs penalties and smuggling, export rebates and processing trade) that answer different questions, is reviewed under administrative-law standards against an authority rather than as a contract, hinges on a technical determination (the heading, the valuation method, the origin finding) settled in administrative decisions and rulings rather than tidy civil judgments, clusters by Customs authority and reviewing court, and lives in Chinese in browse-first databases. How a structured 160M+ judgment-and-ruling corpus — spanning the administrative track — makes classification, valuation, origin, penalty, and rebate precedent findable in English, by claim type, track, governing rule, court, and disposition, with case-number citations to the originals — for cross-border counsel, importers and exporters, customs brokers, and legal AI.

When cargo arrives short or damaged under a bill of lading, an owner and charterer fall out over hire and demurrage, two ships collide, a creditor arrests a vessel to secure a claim, a mortgagee enforces against a ship and competing maritime liens line up for priority, or a shipowner invokes limitation of liability, the matter is a maritime dispute — and in China it is heard not by the ordinary civil courts but by a dedicated system of specialized maritime courts, under a separate Maritime Code and a web of international conventions. Why that precedent is plentiful yet almost unusable without structure: maritime is not one dispute — it splinters into claim types (carriage and bills of lading, charterparties, collision and salvage and general average, ship arrest and maritime liens, ship finance and mortgages, marine insurance) that answer different questions, sits in its own forum and its own code, keeps its decisive steps (arrest, lien priority, limitation) in procedural rulings rather than judgments, turns on standard-form and convention wording, and lives in Chinese in browse-first databases. How a structured 160M+ judgment-and-ruling corpus — rulings included — makes carriage, charterparty, collision, arrest, and lien-priority precedent findable in English, by claim type, governing rule, maritime court, and disposition, with case-number citations to the originals — for cross-border counsel, shipowners, cargo interests, P&I clubs, and legal AI.

When an insurer denies a claim for non-disclosure, an exclusion clause is contested, a liability policy is triggered by a product defect or construction accident, or an insurer that has paid pursues subrogation against the at-fault third party, the matter is an insurance dispute — where large, recurring exposures concentrate. Why that precedent is plentiful yet almost unusable without structure: insurance is not one dispute — it splinters into policy lines (property and cargo, motor, life and health, liability, credit and guarantee, reinsurance) and cross-cutting doctrines (duty of disclosure, the insurer's duty to explain exclusions, subrogation) that answer different questions, turns on doctrine and standard-form wording rather than the signature alone, buries the decisive findings (was the exclusion explained, was the non-disclosure material) in prose, and lives in Chinese in browse-first databases. How a structured 160M+ judgment-and-ruling corpus makes coverage, disclosure, exclusion, and subrogation precedent findable in English — by policy line, clause, doctrine, and disposition, with case-number citations to the originals — for cross-border counsel, insurers, reinsurers, and legal AI.

When an off-plan apartment is delivered late or not at all, a developer defaults and creditors, buyers, and lenders all reach for the same half-built project, a land-use right is transferred but never registered, a bank enforces a mortgage over PRC housing, or co-owners fight over who holds title, the matter is a real estate or property dispute — where the largest cross-border exposures concentrate. Why that precedent is plentiful yet almost unusable without structure: real estate is not one dispute — it splinters into claim types over immovable property (off-plan and commodity-housing sales, developer default, land-use rights, real-property mortgages, leasing and property management, co-ownership and title) that answer different questions, turns on property rights and registration and priority rather than signatures alone, is heavily region- and policy-sensitive, clusters by developer-distress cycle, and lives in Chinese in browse-first databases. How a structured 160M+ judgment-and-ruling corpus makes delivery, mortgage-priority, land-use-right, and title precedent findable in English — by claim type, the right at stake, registration status, region, and year, with case-number citations to the originals — for cross-border counsel, secured lenders, distressed-asset investors, and legal AI.

When a Chinese regulator fines a company, revokes a license, seizes goods at customs, or refuses to disclose the record behind a decision, the remedy is not a commercial lawsuit — it is administrative litigation, a challenge to the legality of a government act, governed by its own Administrative Litigation Law. Why that precedent is plentiful yet almost unusable without structure: suing the state is not one dispute — it splinters into challenges to different government actions (penalty, licensing, customs, land and planning, information disclosure, agency inaction) that answer different questions, is decided on administrative-law standards (legal basis, authority, due process, proportionality) rather than civil liability, produces review dispositions (uphold, annul, confirm-unlawful, order-to-act, remand) rather than win-or-lose, is entangled with a prior administrative-reconsideration step, is published unevenly, and lives in Chinese in browse-first databases. How a structured 160M+ judgment-and-ruling corpus makes penalty, licensing, and customs precedent findable in English — by action type, agency level, disposition, region, and year, with case-number citations to the originals — for cross-border regulatory counsel, compliance teams, and legal AI.

When a Chinese counterparty files for — or is pushed into — bankruptcy, everything a foreign creditor thought it was owed turns on a different body of law: liquidation or reorganization, where an unsecured claim ranks, whether a pre-filing transfer can be clawed back, whether a controller is personally liable, and whether a Chinese court will recognize a foreign proceeding over PRC assets. Why that precedent is plentiful yet almost unusable without structure: insolvency is not one proceeding — it splinters into proceeding types and satellite disputes (liquidation, reorganization, compromise, claim confirmation, avoidance, director and shareholder liability) that answer different questions, lives as much in procedural rulings as in judgments, spans multiple stages, is developing and forum-sensitive at the cross-border frontier, and lives in Chinese in browse-first databases. How a structured 160M+ judgment-and-ruling corpus makes claim-priority, avoidance, and recognition precedent findable in English — by proceeding type, sub-dispute, region, and year, with case-number citations to the originals — for cross-border creditors, restructuring counsel, and legal AI.

When shareholders fall out over who owns the equity, a co-founder never pays in subscribed capital, a minority investor is frozen out of the books, a resolution is pushed through against the articles, or a deadlocked joint venture must be unwound, the claim in China is a company or shareholder dispute — among the most consequential commercial cases the courts hear. Why that precedent is plentiful yet almost unusable without structure: company law is not one question — it splinters into sub-causes (equity transfer, capital contribution, right-to-know, derivative and resolution-validity, dissolution) that answer different questions, was reshaped by the 2024 revised Company Law so holdings must be read against a moving rule, turns on control and validity findings buried in prose, and lives in Chinese in browse-first databases. How a structured 160M+ corpus makes equity, capital, and control precedent findable in English — by sub-cause, applicable-law period, region, and year, with case-number citations to the originals — for cross-border corporate counsel and legal AI.

When a marriage ends, parents fight over a child, a couple cannot agree on who keeps the apartment, or an estate passes to the next generation, the claim in China is a marriage and family dispute — collectively among the highest-volume civil cases the courts hear, divorce alone being one of the single largest civil categories. Why that precedent is plentiful yet almost unusable without structure: family law is not one question — it splinters into sub-causes (divorce, property division, custody, support, inheritance) that answer different questions, turns on facts and judicial discretion buried in prose, varies sharply by region and over time, and lives in Chinese in browse-first databases with uneven, privacy-redacted coverage. How a structured 160M+ corpus makes division, custody, and support precedent findable in English — by sub-cause, region, and year, read against the reasoning and local pattern of the time, with case-number citations to the originals — for cross-border family counsel and legal AI.

When a product injures a consumer, a car hits a pedestrian, a patient is harmed in treatment, or a factory contaminates a waterway, the claim in China is a tort dispute — collectively one of the largest and most varied bodies of civil litigation the courts hear. Why that precedent is plentiful yet almost unusable without structure: tort is not one standard — it splinters into sub-causes (motor-vehicle, personal injury, product liability, medical damage, environmental) that run on different liability bases (fault, fault-presumption, strict), computes damages from region- and year-specific per-capita figures plus a disability grade, and buries the decisive findings in prose, all in Chinese in browse-first databases. How a structured 160M+ corpus makes liability-basis and damages precedent findable in English — by sub-cause, region, and year, read against the rule and local figures in force at the time, with case-number citations to the originals — for cross-border counsel and legal AI.

Every large project in China is a lawsuit waiting to be filed — construction disputes are among the largest and most technically demanding civil matters the courts produce. Why that precedent is plentiful yet almost unusable without structure: it splits into distinct sub-causes (construction-contract, quality, decoration, survey-design), turns on rules that evolved under the Civil Code (contractor priority of payment, payment even on an invalid contract, the black-and-white contract problem), buries the decisive findings inside some of the longest, most fact-heavy judgments in the system, and lives in Chinese in browse-first databases. How a structured 160M+ corpus makes priority, validity, and delay precedent findable in English — by sub-cause, region, and year, read against the rule in force at the time, with case-number citations to the originals — for cross-border counsel and legal AI.

When money changes hands and the deal goes wrong, the litigation is a lending dispute — and in China these are among the highest-volume civil matters the courts hear, private lending alone being one of the single largest civil categories. Why that precedent is plentiful yet almost unusable without structure: it splits into distinct sub-causes (private lending, bank loans, guarantees, financial leasing), turns on an interest-rate cap that changed regime in 2020 (from fixed thresholds to a multiple of the LPR), buries the decisive figures (agreed rate, adjusted rate, guarantee amount) in prose, and lives in Chinese in browse-first databases. How a structured 160M+ corpus makes interest-rate, guarantee, and default precedent findable in English — by sub-cause, region, and year, read against the cap in force at the time, with case-number citations to the originals — for cross-border counsel and legal AI.

Every commercial relationship in China rests on a contract, so when something breaks the dispute is, overwhelmingly, a contract dispute — the largest single category of commercial litigation there is. Why that precedent is plentiful yet almost unusable without structure: it splinters into sub-causes (sale of goods, construction, loan, lease, service), splits across the old Contract Law and the 2021 Civil Code by date, buries the decisive figures (penalty rates, adjusted damages) in prose, and lives in Chinese in browse-first databases. How a structured 160M+ corpus makes breach, liquidated-damages, and construction-payment precedent findable in English — by sub-cause, region, and year, with case-number citations to the originals — for commercial counsel and legal AI.

The legal question a multinational touches most in China isn't antitrust or IP — it's employment, and labor disputes are one of the largest civil-litigation categories there is. Why the precedent that predicts a termination, severance, or non-compete outcome is plentiful yet almost unusable without structure: it runs across two stages (mandatory arbitration, then court), turns on locally variable average-wage numbers that change by city and year, buries the decisive figures in prose, and lives in Chinese in browse-first databases. How a structured 160M+ corpus makes it findable in English — by cause of action, region, and year, with case-number citations to the originals — for employment counsel and legal AI.

"China data compliance" used to mean a short memo and a shrug. With the DSL and PIPL both in force since late 2021, it's now one of the world's most consequential data regimes — rectifications, fines, a hard cross-border-transfer regime, and a fast-growing body of court litigation. Where the enforcement record lives across the CAC administrative track and PIPL judicial track, why it's hard to assemble (young, two-system, reasoning-heavy, Chinese-language), and how a structured 160M+ corpus makes PIPL and cross-border precedent searchable in English for privacy counsel and legal AI.

A Western legal AI team that already shipped US and EU coverage treats China as "one more jurisdiction" — and that instinct stalls almost every China-coverage project. The five assumptions imported from common-law case law that all break on PRC data: scraping is the core task, precedent is a citation graph, machine translation preserves meaning, published equals complete, and structuring is cheap. Why each fails, and what 160M+ Chinese judgments actually require to become a usable AI corpus.

Separate Chinese source text, English query handling, summaries, structured fields, REST and MCP access, and full translation. A practical SinoVerdict buyer acceptance guide.

Most "add China coverage" projects die when a team drops a pile of judgments into a vector DB and watches it return look-alike cases decided the opposite way. A reference architecture for the five stages that actually make retrieval work over 160M+ judgments — ETL, section-aware chunking, hybrid lexical-plus-vector indexing, filtered retrieval with reranking, and the citation and eval layers — plus a build-vs-license split, layer by layer, for ML and platform engineers.

Ask an AI to list the legal MCP servers and you get CourtListener, FOLIO, legal-mcp — all US/EU, every one stopping at the Chinese border. What an MCP server for PRC case law does, why agentic legal AI needs one over a raw API, how to connect Claude Desktop / Cursor / ChatGPT, and where SinoVerdict's 160M+ judgment MCP server fills the China-shaped hole in your tool list.

"Use CAIL2018" is the right answer to a research question and the wrong one to a product question. What the academic Chinese legal datasets — CAIL2018, LeCaRD/LeCaRDv2, CJO scrapes — are genuinely good for, the five places they break in production (frozen, criminal-only, Chinese-only, unstructured, research-licensed), and what a current, 160M+, English-indexed, license-ready corpus changes for legal AI vendors.

Where do you actually buy Chinese case law for a legal AI product? A map of the five supply categories — official channels, domestic aggregators (PKULaw, Wolters Kluwer China), Western incumbents, raw scrapers, and AI-native structured licensors — scored on coverage, AI-readiness, and license-to-ingest for the vendor's job, not the lawyer's. Most obvious names are excellent lawyer-sources and the wrong shape for a pipeline.

For twenty years, Chinese case law was searchable only in Chinese — or in a few thousand hand-translated cases. Why the language wall stood, why neither hand nor raw machine translation knocked it down, and the four capabilities — structured corpus, bilingual taxonomy, English query, and citation grounding — that finally make English-native research over 160M+ judgments possible for legal AI teams and China-practice lawyers.

China has published 160M+ judgments — but raw volume is a liability, not an asset. The structuring pipeline that turns a heterogeneous document dump into an AI-ready corpus: normalization, field extraction, de-duplication, cause-of-action taxonomy, citation grounding, and a cross-lingual layer — plus the schema and the questions to demand from any vendor.

China went from no Patent Law in 1984 to the world's largest IP docket — half a million matters a year, a dedicated SPC IP Court, punitive damages. What four decades of judgments reveal about enforcement, why that case law is hard to assemble across five rights, and how a structured corpus makes patent, trademark, and trade-secret precedent searchable for IP counsel and legal AI.

China is now one of the world's most active antitrust jurisdictions — multi-billion-yuan fines, a 2022 AML amendment, and a fast-growing body of court litigation. Where the precedent lives across the SAMR and judicial tracks, why it's hard to assemble, and how a structured corpus makes it searchable for competition counsel and legal AI.

Will a Chinese court enforce a foreign arbitral award? The doctrine—New York Convention plus the SPC reporting system—is reassuring; the precedent that proves it is rare, scattered, and Chinese-language. How enforcement actually works, why the rulings are hard to find, and how a structured corpus makes them searchable.

Legal AI vendors don't need a one-time dataset — they need an ongoing data partnership. How bulk corpus, REST API, and incremental sync compose to power a China feature, the licensing structure behind them, and what to ask a data partner before you sign.

"Add China" is a data-infrastructure problem, not a content one. The six-layer stack you actually build to ground a legal AI product in Chinese case law — corpus, ingestion, structure, retrieval, citations, evals — and which layers to buy instead of build.

Scraping Chinese judgments looks free until you price in the coverage holes, the legal exposure, and the forever-maintenance. An honest build-vs-buy breakdown for legal AI teams — and when licensing wins.

LexisNexis lists ~3,000 translated Chinese documents; vLex omits mainland primary law. Both were built for human research, not retrieval at scale — here's the volume, format, and license gap legal AI teams hit, and how to close it.

A Chinese court-data licensing checklist for internal use, end-user display, redistribution, model training and translation, with separate technical acceptance and rights review.

How PRC judgments are structured — case numbers, cause-of-action taxonomy, reasoning sections — and how to integrate a Chinese case law API for RAG and evals.

China published 19.2M judgments in 2020, ~5.1M in 2023, ~11M in 2025. Get the full availability curve and how to vet Chinese case law data before licensing.