Case law coverage

China Arbitration Case Law Data: 186,000 Court Records, and the Thin Foreign-Award Slice

"China arbitration data" is not one dataset, and the difference between its slices is two orders of magnitude. A buyer who asks for arbitration coverage may mean set-aside applications, which are abundant, or recognition and enforcement of foreign awards, which is a narrow practice with a few hundred records. In our own corpus the union of twenty arbitration-related causes of action returns 185,793 court records; a single cause inside that union, recognition and enforcement of foreign arbitral awards, returns 125.

This walkthrough sets out what the corpus actually contains, where the volume sits, which fields mislead, and how to write a pilot that can fail for the right reasons. All counts were read live on 22 October 2026 against an index of 160,291,678 records with benchmark entries excluded. Every figure is a candidate pool of court records. None of them is a count of arbitrations filed, a count of awards issued, or a success rate.

1. Four slices that get sold as one product

Chinese court documents about arbitration cluster for procedural reasons. An arbitration commission issues the award; a court handles what happens next, and each of those handling steps has its own cause of action. That produces four natural slices, and they are not interchangeable.

A vendor that answers "how much China arbitration data do you have" with one number is usually quoting the first slice and selling the fourth. Ask which causes are included, and ask for the per-cause breakdown before you agree to anything. The earlier note on foreign arbitral award precedent makes the same point from the research side.

2. What the corpus returns, cause by cause

Twenty cause labels in our index mention arbitration, recognition or foreign judgments. Their union returns 185,793 records. Three groups inside it explain almost all of the volume.

The domestic group is dominated by two labels: 申请撤销仲裁裁决 (applications to set aside an arbitral award) with 84,338 records and 国内非涉外仲裁裁决 (domestic, non-foreign-related arbitration awards) with 57,301. Validity of the arbitration agreement adds 12,152, property preservation inside arbitration proceedings adds 8,099, and general arbitration procedure cases add 3,123. The remainder are small: 227 records for arbitration conduct, 22 for evidence preservation and 16 for pre-arbitration preservation.

The recognition group is nearly one label: 申请承认与执行法院判决、仲裁裁决案件 returns 17,338 of its 18,045 records. That label deliberately mixes court judgments with arbitral awards, so it cannot be used as an award counter without splitting the rows. Two separate labels cover recognition of foreign court judgments, at 349 and 336 records, and a further 22 records sit under a shorter spelling of the same idea.

3. The foreign-award slice is thin, and one batch dominates it

This is the number that matters most to cross-border counsel, and it is small enough to quote exactly.

Two readings follow. First, the group total is inflated by one label: Macau awards account for 1,428 of 2,470 records, and their yearly distribution is concentrated — the whole group moves from 238 records in 2024 to 1,835 in 2025, and that jump is a batch arrival rather than a fivefold rise in filings. Second, the genuinely treaty-facing labels are tiny: 125 records for recognition and enforcement of foreign arbitral awards and 32 under a shorter label for the same objects.

A product that answers "find me Chinese court decisions on enforcing foreign arbitral awards" is therefore working with a small, high-value set. That is a legitimate design, but it must be described as what it is: a few hundred documents nation-wide, reviewed individually, where each new ingestion matters. Promising "comprehensive China arbitration coverage" and delivering this slice is the most common scope failure we see. See how ingest batches distort time series for the general pattern.

4. Records are not cases, and the gap is not constant

The union returns 185,793 records but only 171,340 case numbers that end in the standard character, and about 148,668 distinct raw case numbers. The gap between those columns is the part that quietly breaks dashboards.

Multiplicity is not stable: it is about 1.01 in the early years and 1.85 in 2019. Enforcement and preservation matters produce several documents per case number — an application, a ruling and a follow-up order — while the earlier population is closer to one judgment per case. The case-number estimate uses raw strings, so bracketed or spaced variants count separately; it is a diagnostic, not a de-duplicated case count. A chart built on records will show a growth curve that a chart built on distinct case numbers does not reproduce.

5. Field-level traps that survive a visual check

Three fields behave badly in this vertical, and each one produces a confident wrong answer.

The province field is a court-location label, not the parties' address, and the non-contentious procedure population was ingested with weaker mapping — hence the blank share. The procedure field is the classic case of an unfinished crosswalk: numeric codes, Chinese labels and one cause-name string sitting in the same column. Treat each as its own field with its own dictionary and its own coverage figure, and never let two fields be filtered as if they were the same taxonomy. The cause-label era effect documented in the case-type filter note applies here too, because new labels enter the vocabulary years after the practice begins.

6. Text evidence: what the documents actually say

Cause labels tell you how a court docketed a matter. Word counts tell you what the text discusses, and for arbitration the text is unusually quiet about its own governing law.

The New York Convention is named in 188 records across the entire corpus and in zero document titles. Within the arbitration union the short form appears 72 times and the full treaty name 77 times. That is not evidence that Chinese courts ignore the convention; it is evidence that the convention is usually addressed by article number or by paraphrase rather than by name. A retrieval product that only matches 纽约公约 will miss most of the reasoning, and a product that reports "no treaty discussion found" from a phrase count has made a measurement error, not a legal finding.

Two method notes keep these counts honest. Phrase behaviour depends on the analyzer: a phrase query against the wrong analyzer can return a false zero, so our counts take the union of two analyzer modes rather than a single match — see the analyzer mismatch note. And a phrase hit is a word occurrence, not a holding: 不予执行 in 4,619 records includes applications, submissions and quotations, not only grants. Field presence is also not the same as readable reasoning, which is why we publish the full-text coverage guide alongside any count.

7. A data model for arbitration records

Because this material is procedural, the entity model matters more than the text volume. Five objects should be separate rows.

Keep the court's own case number as a stable identifier. In our corpus 92 per cent of records across the whole collection carry the court's case number, and the arbitration union is close to that level, so the field is usable — but it must be stored in both its printed and normalised forms, and never inferred. The wider metadata census documents how inconsistent these fields are across the corpus.

8. A pilot that can fail for the right reasons

Because the foreign-facing slice is small, a pilot on it should be reviewed document by document. Build the acceptance set before the pipeline and include a refusal, a partial grant, a case where the court mentions the convention by article only, and a matter with several documents under one case number.

Measure retrieval and extraction separately. A record that exists but is never retrieved is a collection failure; a retrieved record whose outcome is invented is an extraction failure. The general handover checklist in our delivery acceptance checks applies directly, and the mass-filing caution in the cluster note explains why a small sample in this vertical must be read individually rather than extrapolated.

9. Commercial scope and language are separate questions

Public accessibility and permitted commercial use are different tests, and award-related records are among the most fact-specific documents in the corpus. Record where each item came from, and obtain review for storage, display, redistribution and model training rather than inferring a grant from a successful retrieval; our licensing guide treats these as rights questions to confirm.

SinoVerdict's 160M+ figure describes court-data records and implies nothing about the number of arbitrations, awards or commercial commissions in China. Request a scoped trial API key and corpus coverage report, and discuss bulk, REST and MCP delivery against an agreed scope through our delivery overview. Test language behaviour on the endpoint you will actually call: Chinese source text, an English interface, English query support and a translated corpus are four capabilities, as the English-access capability matrix sets out.

Frequently asked questions