Home → Insights → Foreign exchange enforcement data

China Foreign Exchange Enforcement Data: SAFE Disclosures and Court Judgments

A search for China foreign exchange enforcement data can lead to an agency penalty notice, a court judgment, a policy circular or a report describing selected violations. All four can help a compliance team. They do not describe the same population, and none should silently stand in for the others.

For an overseas legal AI vendor, the procurement question is practical: which source supplies the administrative decision, which supplies judicial reasoning, and which fields let a reviewer connect the two? This guide maps official SAFE discovery routes, reports three live searches of SinoVerdict’s court-record index, and turns the boundaries into an evaluation specification. It does not estimate the total number of foreign exchange violations in China.

Start with the document your product needs

“Enforcement case” is too broad for a purchase order. An assistant explaining a regulator’s stated reason for a fine needs the agency decision. An assistant researching a challenge to that decision needs the judicial document. A product explaining a criminal allegation needs a criminal track. A business dispute mentioning a payment abroad may require none of those labels.

Store the track before extracting an amount or generating a summary. Otherwise a model can confidently call a contractual payment a penalty, or present an allegation as a finding. Our administrative litigation dataset guide covers the broader court-document distinction.

Where SAFE publishes the discovery routes

The SAFE enforcement disclosure directory links to administrative penalty information. Its penalty-query landing page embeds a query surface. Follow the actual query and its scope notice rather than assuming the surrounding landing page contains the records.

The national query surface provides an organization-code input and describes recent central-office disclosure coverage, with branch information located on branch sites. That scope notice matters more than the official domain name: a central-office window is not a nationwide historical register. An empty extracted table is also not proof that the agency issued no penalties. It can reflect the requested scope, rendering or query state.

Branch context changes the view. The Beijing query surface showed dated entries and decision references when checked for this article. This demonstrates a usable discovery route; it does not establish nationwide completeness or a guaranteed retention period. Keep the branch parameter and observation date with each source inventory entry.

A buyer seeking historical coverage should request a branch-by-year availability matrix. The matrix needs separate states for observed records, a genuine empty result, an unavailable page and a page that could not be evaluated. Collapsing the last three into zero manufactures a coverage claim the source never made.

Keep the rule library alongside the decision data

SAFE’s inspection and application-of-law directory supplies a separate discovery route for governing materials. For example, its 2022 administrative-penalty announcement identifies an effective date of June 1, 2022 and replaces the preceding 2020 measures. That is a versioning anchor for the rule collection, not evidence that a particular court record applied those measures.

The same directory separately lists the SPC and SPP interpretation on illegal payment settlement and foreign exchange trading. Its subject explains why a search vocabulary may retrieve criminal material as well as administrative material. We cite the listing to identify the source family; this article does not interpret offense thresholds or advise whether particular conduct is lawful.

For a retrieval product, preserve original Chinese titles, issuing authorities, publication dates, effective-date evidence and version status as distinct fields. An English topic label can help discovery, but should not overwrite the official title. Current directory placement alone cannot establish which version governed conduct years earlier.

What three live court-record searches actually returned

On October 9, 2026 we ran three queries against SinoVerdict’s judgment-record index. Each query takes the union of phrase searches using ik_smart and ik_max_word. Completed responses had no shard failures. We used the union because Chinese tokenization can change recall; the title search itself illustrates a small difference.

These are three different candidate pools. The first uses a shorter agency expression in the title; the second uses a central-agency name in the body; the third uses a conduct expression. They overlap, so do not add the counts. Nor does their relative size measure how often regulated entities challenge SAFE. No procedure filter, distinct-case count or semantic relevance audit is included in this measurement.

Body matches may be citations to a rule, background allegations or references to another proceeding. A phrase match tests analyzed token positions, not a guaranteed literal substring and not a legal conclusion. Equal counts from two analyzers do not validate the relevance of every record. See our analyzer-mismatch audit for the retrieval issue.

The downloadable query definitions and aggregate responses make this snapshot reproducible. They contain no individual case text. Future counts may change with the index. The observation date belongs in any quotation or coverage comparison.

Specify a topic slice before asking for a price

A useful request names agencies and branch aliases, target actions, procedural tracks, regions and a date range. It also says whether the task is case research, a benchmark or a training corpus. Those objectives need different acceptance evidence. A broad word match can help discover vocabulary; it is not a finished delivery specification.

For example, an evaluation scope could target court review of foreign exchange penalties and keep criminal trading allegations in a separate set. Require a reviewer to identify the government action being discussed and the passage that supports the label. Put ambiguous matches in a review bucket. Excluding an uncertain record is different from asserting that it is irrelevant.

Do not start with a desired record total and broaden the query until it is reached. Start with reviewed examples representing the intended task, record the chosen inclusion rules, then measure retrieval against those rules. A smaller inspected set can be more useful for a pilot than a larger unlabeled pool.

Connect decisions using evidence, not names alone

An agency decision number and a court case number identify different documents. A court record can mention several decisions, and one regulatory action may appear in more than one judicial document. Keep both original strings and the passage supporting their relationship. Avoid a database design that forces a one-to-one match before you know the relationship.

Entity names need equal care. A branch, a bank headquarters and a named employee can appear in the same record with different roles. Match the role and source coordinates before proposing an entity link. Do not expand masked individual names or infer an identity from initials.

Keep proposed links separate from reviewer-confirmed links. When supporting text is missing, the right result is an unresolved relationship. The interface should display that state clearly instead of producing an authoritative-looking chain from fuzzy name similarity.

Ask for these fields in a buyer’s coverage report

These are evaluation requirements, not a claim that every field is available in every SinoVerdict record. Require completeness and accuracy to be reported separately. A derived field may be filled consistently while still being wrong; our derived metadata audit explains why that distinction matters.

Run an evaluation that can fail visibly

Choose the intended tracks and date strata before sampling. Record the sampling frame, random seed, sample size and handling of duplicates. Inspect the selected passages rather than only the result titles. A relevant title with missing body text is a different failure from a fully readable but irrelevant document.

Test questions should distinguish the agency’s allegation from the court’s own finding, and the challenged decision from a separate decision quoted in the reasons. Include ambiguous records and missing-evidence cases. An assistant that declines to infer an unsupported link is behaving better than one that always returns a fluent answer.

For each failed answer, retain the question, retrieved record IDs, available evidence, English output and reviewer correction. That creates a repairable error list. A single aggregate score conceals whether the defect is coverage, retrieval, translation or evidence handling.

No sampling experiment or relevance percentage is reported here. The live counts are a discovery snapshot. An enterprise pilot should establish those additional measures for its own slice before treating the candidate records as a validated dataset.

What a court corpus cannot tell you

Judicial material reflects matters that produced publicly available court records and were captured in the index. It cannot by itself establish how many SAFE decisions were issued, how many were challenged, or how often a regulated entity succeeded. Those rates require a defined denominator and reliable links across source tracks.

Likewise, absence from a query can mean different things: no captured record, different vocabulary, missing text, a field defect or an unavailable source. Report the observed limitation. Do not turn a retrieval miss into a statement that a company has never faced foreign exchange enforcement.

SinoVerdict’s public 160M+ record description concerns its broader court-record corpus. It is not a claim of 160 million distinct foreign exchange cases, complete readable text or universal official-source URLs. Court case numbers and other available coordinates support checking; they are not the same as a working link to an original judgment.

Move from discovery to an enterprise delivery

Use the official directories to define the regulatory source families, then test court retrieval on a reviewed scope. Ask for a coverage report that states which tracks were evaluated, which were excluded, what text is available and what remains uncertain. An example record should demonstrate the intended product workflow rather than merely show that an API returned JSON.

SinoVerdict offers enterprise court-data access through bulk delivery, REST API and MCP workflows. For this topic, begin with a trial API key and scoped corpus coverage report; use the delivery overview to discuss integration. Confirm the actual slice, available fields, English output and permitted uses in the evaluation and agreement. Public access to an agency page should not be treated as a blanket authorization for every downstream use.

Frequently asked questions