Regulatory data sources

China Work-Safety Enforcement Data: Emergency Management Penalties and Court Records

Chinese work-safety enforcement produces two different kinds of document, and only one of them is a judgment. The Ministry of Emergency Management and its provincial and municipal departments publish administrative penalty decisions and accident investigation reports. Courts publish criminal and civil judgments about the same events. A product that answers "how is work safety enforced in China" needs the publication series; a product that answers "how have these cases been decided" needs the judicial record. Merging them into one figure produces a number that describes neither.

This guide sets out where the official publications sit, shows what our own court corpus does and does not contain for this domain using live aggregate checks dated October 21, 2026, and proposes a data model and acceptance plan for overseas teams. It does not publish an enforcement rate, an accident rate or a risk score, and it is not a statement about any institution or any incident.

1. Two tracks, two questions, two formats

The distinction is not academic. An administrative penalty decision carries a decision number, the sanctioned party, a description of the violation, the measure imposed, the deciding authority and a date. An accident investigation report is a narrative document that establishes what happened, who was responsible and what is recommended. A criminal judgment names defendants, applies a provision of the Criminal Law and records a sentence. Each has its own identifiers, its own issuing body and its own publication rhythm.

Three buyer questions therefore need three different datasets. "Which entities were sanctioned for safety violations last year" needs the regulator's disclosure series. "How are major accidents investigated and what is recommended" needs accident investigation reports, which are usually longer documents with their own numbering. "What sentences follow fatal construction accidents" needs the judicial record, filtered by criminal cause rather than by regulator name.

The failure mode is always the same: a team collects the judicial record because it is available as a corpus, then reports its counts as enforcement statistics. Court records are the tail of a much longer administrative process, and in this domain the tail is thin. The general distinction between a penalty publication and a judicial document is set out in our administrative penalty source guide.

2. Where the official publications sit

The Ministry of Emergency Management publishes through its own site. Three entry points were checked on October 21, 2026 and each returned HTTP 200: the ministry home page, the administrative law-enforcement disclosure section (行政执法公示), and the service and inquiry section (查询服务). The ministry's own inquiry page lists the discovery routes a data team should start from: incident and disaster information (灾害事故信息), incident and disaster investigation (事故及灾害查处) and administrative law-enforcement disclosure (行政执法公示).

Two practical notes follow. First, the page formats differ: the same violation may be described in a table on one site and in a narrative paragraph on another, so a single parser cannot be assumed to cover both. Second, the date roles differ: the date an incident occurred, the date a decision was taken, the date a report was published and the date a page was retrieved are four separate values, and substituting one for another is the most common defect in this domain.

Because the ministry's disclosure pages are paginated by publication rather than by decision, a collection plan needs a stable ingestion order and a recorded retrieval date per item. A later re-read is not evidence that an item changed; only a changed page hash with two retrieval dates supports that claim.

3. What our court corpus actually contains

Our corpus holds 160,284,678 records with benchmark entries excluded, checked live on October 21, 2026. Filtering to the criminal causes that carry safety meaning, a union of eight causes returns 9,346 records covering 8,885 distinct case numbers under a case-number pattern check. The distribution across causes is informative because it shows how concentrated the domain is.

These are candidate retrieval pools from one corpus, not a national count of accidents, prosecutions or sanctions. Distinct case numbers are lower than record counts, which is expected: the same case can appear in more than one collection, and the same incident can generate several judgments. Any published figure should state which of the two it is.

A second structural point concerns the cause field itself. The 2021 amendment renamed one offence, so both the old and the new wording appear in the data, and a filter built on only one of them silently loses records. This is the same class of defect as analyzer mismatch in phrase queries, described in our phrase-query note: the query returns a plausible number, and the number is wrong.

4. The year curve is a coverage curve first

Plotted by year, the safety-cause pool rises from a handful of records in the 1990s to 1,799 records in 2015 and then falls to 593 in 2018, 170 in 2019 and 97 in 2020. Our slice contains a single record dated 2021, 75 in 2022, 32 in 2023 and 33 in 2024, then 261 records dated 2025 arriving through a distinctly newer collection slice. The corpus as a whole shows the same shape, with published volume contracting sharply after 2018 and 2023–2025 far below the 2022 figure.

The temptation is to read the fall as improving safety. It cannot be read that way. A single record in 2021 is a statement about our slice, not about the year, and the same publication contraction appears across every domain we measure. Presenting this curve as an enforcement trend without the coverage curve beside it would mislead a reader in the most quotable possible way. The general treatment of this problem is in our coverage and recency note.

What the curve can support is a scoping decision. If a buyer's questions are about 2013–2017 practice, this pool is dense enough to sample and review. If the questions are about 2023 onwards, the judicial pool is thin, and the administrative publication series becomes the only usable source — which is exactly why the two tracks must be collected separately from the start.

5. Geography, procedure and collection mix

Province values in the safety-cause pool are led by Henan (1,319), Jiangsu (1,018), Shandong (593), Zhejiang (579) and Guangdong (447), with 691 records carrying no province value at all. Province is a field of the deciding court, so it measures where the case was decided rather than where an incident occurred, and a company operating across provinces will appear in several buckets. Unmapped values must be reported, not silently dropped, or every rate computed from the remainder inherits a hidden denominator.

The procedure field is similarly uneven: 7,172 records carry the first-instance code, 752 a second code, 667 a third, and a minority carry a literal label such as 一审 or 刑事一审. Mixed code-and-label vocabularies are normal in this corpus and mean that any procedure filter needs an explicit mapping table with the original value retained.

The collection mix explains the recency behaviour directly. In this safety pool, 8,608 records come from the main historical slice while 261 come from a newer annual slice, 177 and 159 from two 2019 slices, 95 from a 2023 slice and 33 from a 2024 slice. Records coming from different collection batches can carry different field quality and different anonymisation behaviour, so the batch identifier belongs in the data model, not only in the ingestion log.

6. The regulator is almost absent from the judicial text

The clearest evidence that the two tracks are separate comes from phrase checks against the whole corpus. The term 应急管理 (emergency management) appears in 7,665 records out of 160,284,678. The phrase 生产安全事故 (production safety accident) appears in 19,672 records, 事故调查报告 (accident investigation report) in 42,282, and 安全生产 (work safety) in 300,728. A generic phrase such as 行政处罚决定书 (administrative penalty decision) appears in 2,057,954 records.

The decisive number is the overlap. Of the 9,346 records in the safety-cause pool, only 21 contain 应急管理, while 494 contain a penalty phrase and 2,989 contain the accident investigation report phrase. In other words, the regulator that issues the penalties is named in about two in a thousand of the criminal judgments about the events it regulates. A team that tries to reconstruct the enforcement record from court text will produce a sample so small that no rate derived from it means anything — and, worse, it will look like a complete answer.

Two method notes make those figures reproducible. Phrase behaviour depends on the analyzer, so our counts take the union of two analyzer modes rather than a single match, because a phrase query against the wrong analyzer can return a false zero. And a co-occurrence is not a role: a judgment containing the accident investigation report phrase may quote the report as evidence, summarise it, or dispute it. Sample and describe the document role before publishing any rate.

7. A data model that keeps the tracks apart

Model the administrative and judicial objects as separate entities with their own identifiers, then link them only through reviewed evidence. The following fields are the minimum for the administrative side.

Keep the original Chinese text and any English label side by side, and store a text span or row locator for every extracted field so a reviewer can verify a value without re-reading an entire publication. Version the parser so that a schema change is distinguishable from a source correction.

Person data deserves its own policy. Penalty decisions and accident reports routinely name responsible individuals, sometimes with partial anonymisation. What is stored, how long it is retained, how it is displayed and whether it may be used for training are rights questions to be confirmed rather than defaults to be assumed; our licensing guide treats them that way.

8. Acceptance tests that can fail for the right reasons

Build the acceptance set before the pipeline, and include the cases that break naive implementations: a publication covering several decisions against several parties; a report whose publication date differs from the incident date; a company penalised at both provincial and municipal level; a judgment that quotes an investigation report without being about the accident; and a cause label that changed wording after 2021.

Measure retrieval and extraction separately. A decision that exists but is never retrieved is a collection failure; a retrieved decision with an invented fine amount is an extraction failure. Report the denominator, the labelling method and the reviewed error rate for each. For the wider handover process, our delivery acceptance checks apply directly to the judicial component.

9. Commercial scope and language are separate questions

Public accessibility and permitted commercial use are different tests. Regulatory publications and a licensed judicial collection can carry different permissions, and person-level records deserve the strictest review of the three. Record where each item came from and obtain review for storage, display, redistribution, translation and model training rather than inferring a grant from a successful download.

SinoVerdict's 160M+ figure describes court-data records; it is not an administrative penalty register and implies nothing about administrative coverage. Request a scoped trial API key and corpus coverage report for the judicial component, 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, generated English summaries and a fully translated corpus are five different capabilities, as set out in the English-access capability matrix. A work-safety pilot should run its own known Chinese and English queries rather than borrowing a general statement about coverage.

Frequently asked questions