Regulatory data sources

China Tax Enforcement Data: STA Penalty Notices and Court Records

China tax enforcement data lives in the tax authority's own publication series. The court corpus is dominated by invoice-related criminal prosecutions, which are a different object from an administrative tax assessment, and the two are almost never the same event. A dataset that merges them produces a number that cannot survive an audit: the largest tax-related pool in a judgment corpus records who was prosecuted for the paper trail, not what the authority enforced.

This guide maps the official starting points as of October 25, 2026, sets out what our own corpus actually contains under tax and invoice cause labels, and proposes a data model and acceptance-test list for overseas legal AI teams. It is informational only and is not legal advice. Every court-side number below is an aggregate query result over our own collection of 160,284,678 records (benchmark rows excluded), not an enforcement statistic: a corpus record is a document, not a penalty decision, not a distinct case and not a finding of liability.

1. Three tax questions that sound alike and need different evidence

"How many enforcement actions did the tax authority take last year?" needs the authority's publication series with decision dates, parties and the amounts assessed. "Which conduct gets penalised, and how severe?" needs violation typing plus the underlying rule text. "Is this counterparty currently under investigation?" needs today's official status, which no historical publication list can supply.

There is a fourth request that arrives disguised as data work: "show me the court record that proves the tax decision." A judgment can quote a penalty decision, reproduce an assessment, or review the decision on appeal. It can also mention a tax bureau because the bureau is a creditor in a bankruptcy, a party in a construction dispute, or a lienholder in a vehicle case. In our corpus, the phrase 税务局 (tax bureau) appears in 300,469 documents and only 5,708 of them sit under the general administrative cause labels at all — the rest are contract, loan, construction and labour disputes in which the tax authority is incidental.

Write the target answer before collecting a first row, then decide which track owns it. Our administrative-penalty, banking and securities source guides cover the same split in other regulated sectors.

2. Where the official tax publications sit

The anchor is the State Taxation Administration (国家税务总局, STA), which publishes policy, disclosures and enforcement material through its own portal, with provincial and municipal bureaus publishing locally. We probed the portal on October 25, 2026 and recorded the exact responses below.

Two conclusions follow. The authority's own site is reachable but is not a bulk API: listings are rendered, paginated and organised by disclosure category, so extraction needs rendering or an approved interface rather than a static GET. And the aggregator layer that appears to solve this is not reliably available to unattended clients — 信用中国 returned an edge block, which is a measurement about our client rather than about their coverage.

3. The court corpus is an invoice-crime corpus, not a tax-assessment register

Filtering our collection by cause labels that contain either 税 or 发票 returns 38 tax-related label families. Grouping them, the largest object by a wide margin is the invoice-offence family: 37,714 records across 23 cause labels, of which 37,697 carry a case number ending in 号. Tax-payment offences proper are two orders of magnitude smaller: 1,050 records across 11 labels, 1,049 numbered. The two label sets are disjoint, so the combined tax-crime pool is 38,764 records.

Read that table as a property of how Chinese courts label criminal files, not as a statement about national tax enforcement. The offence at the top is the one that accompanies a fabricated value-added tax invoice, an offence introduced into the Criminal Law to protect the invoice system, and the volume reflects prosecutorial practice in a single high-frequency crime type. Tax evasion proper — 逃税罪 — sits at 793 records in the same collection, which is what makes the merged number wrong: an analyst who sums "tax cases" and reports 38,764 enforcement actions has counted mostly prosecutions for paper instruments.

For contrast, 虚开增值税专用发票 appears in the body text of 52,669 documents, 逃税 in 20,037 and 追缴税款 in only 1,799. Phrase counts exceed labelled-cause counts because the phrase is discussed in judgments whose cause label is unrelated, including contract disputes over who bore the tax.

4. One offence, two labels: a data-quality trap you can measure

Inside the invoice family, the same offence is recorded under two labels that differ only by the closing 罪 (crime) character. Both appear as live cause values, and the split is not noise on the margin:

Eleven of the twenty-three invoice labels in our data are the crime-suffixed form and twelve are not. A pipeline that filters on one form loses between 20% and 94% of a given offence depending on which form it picked: choosing only the crime-suffixed 虚开增值税专用发票… 罪 misses 5,372 records, and choosing only 非法出售用于骗取出口退税、抵扣税款发票 (14) while ignoring the crime-suffixed form (12) would be wrong in the other direction. The correct handling is not to normalise silently: keep the original label, add a normalised code alongside it, and report the label-level breakdown. Our metadata field census documents the same class of problem across other fields, and the raw counts behind this page ship as a JSON snapshot.

5. The year profile has three artefacts, and only one of them is about enforcement

Bucketing the invoice-family pool by the corpus year field produces a shape that is easy to misread.

Three things are happening at once. The 2014–2017 plateau is the period when this offence type was both prosecuted at scale and published at scale in the corpus. The 2019–2020 trough (519 and 297 records) coincides with the nationwide collapse in judgment publication, which affects every cause in the collection and is not specific to tax — see our coverage and recency analysis. The 2021–2022 recovery is a partial return of publication volume, not a doubling of prosecutions: the annual total never returns to the 2015 peak, and 2023 drops again to 775 before 2024 recovers to 1,560.

The fourth column shows why the year field is not an enforcement date: documents citing 国家税务总局 rise to 11,411 in 2020 and 14,130 in 2022 for reasons unrelated to enforcement output. A regulator's name in court text measures judicial vocabulary.

6. Administrative tax litigation is a separate and much smaller track

When the actual question is judicial review of a tax decision, the right pool is administrative litigation with a tax cause label. In our collection that pool is 423 records, every one of them carrying a numbered case file — roughly 1.1% the size of the invoice-crime pool. Its distribution is not stable either: 31 records in 2018, 58 in 2019, 66 in 2020, 105 in 2021, then a single record in 2022, 9 in 2023, 137 in 2024 and 16 in 2025. A single-record year sitting between 105 and 137 should be disclosed as a gap rather than smoothed.

Broadening to the general administrative cause labels and intersecting with the body phrase 税务局 gives a larger and more even pool of 5,708 records, dominated by 行政复议 (2,021), 行政处罚 (1,693) and 行政强制 (1,532). Even that pool concentrates in a few provinces: 浙江 108 and 山东 75 lead the narrow tax-cause pool, while in the wider 税务局 phrase pool 浙江 (31,675) and 广东 (22,507) lead, with 26,713 documents carrying no province value at all.

7. Which document types actually appear inside judgments

The tax administration issues documents with distinct names, and a dataset built on court text can count how often each name is cited. Measured as body phrases across the whole collection, the practical ordering is: 税务局 300,469, 税务机关 150,329, 国家税务总局 70,032, 税务处理决定书 9,008, 税务行政处罚决定书 7,123 and 税务稽查 6,631.

The order matters more than the values: the institution's generic name appears far more often than the instruments it issues. A case that cites a 税务处理决定书 is showing the administrative decision as evidence — the thin, checkable slice a legal AI product can support.

8. Two fields that break naive joins

The procedure field mixes numeric codes with Chinese labels. In the invoice-family pool the top values are the bare codes 1 (21,249), 4 (5,361) and 2 (2,229), alongside the labels 刑罚与执行变更 (3,107), 刑事一审 (2,498) and 一审 (595), so the same offence family appears as first instance in four encodings. Group by this field without a crosswalk and the stage distribution is meaningless — see our procedure code crosswalk. The province field is uneven and sometimes empty: the pool concentrates in 上海 (8,282), 江苏 (5,273), 山东 (3,134), 浙江 (3,111) and 广东 (1,856), with 2,158 records carrying no province value. That shape can reflect prosecution intensity, centralised case assignment or batch publication by one court, and the field alone cannot tell them apart.

9. Acceptance tests before you sign

The publication series and the judicial corpus fail in different ways, so the acceptance test asks different questions of each. Each item is designed so that a partial delivery fails it.

The last item is the one most often skipped: one successful run proves the pipeline works on one snapshot, and only a second run shows whether the source is stable. Our delivery acceptance checks and licensing guide cover the wider process.

10. Commercial scope and language are separate questions

The commercial decision is what the licence permits: redistribution, derivative works, model training, internal research and external display are five different permissions, and a public disclosure page settles none of them. The language decision is how much of the Chinese source is carried through, and promising English-language tax reasoning means translating free-text violation descriptions — a different quality problem from translating structured fields. Our licensing guide and English-native research note cover that track separately.

Frequently asked questions