Home → Insights → Personal information public interest litigation data

China Personal Information Protection Public Interest Litigation: Sources and Court Data

A China personal information protection public interest litigation dataset needs more than a search for privacy terminology. The relevant material can be a prosecutor’s account of supervisory work, a civil judgment, a criminal judgment with a related civil claim, or a court document that merely mentions another proceeding. Those records answer different questions.

For a legal AI vendor, this distinction determines what the product can support. A source map can explain how a matter reached a remedy. A judicial corpus can support research on recorded court reasoning. Neither automatically supplies all investigations, administrative rectifications or people affected. This guide identifies official discovery sources, reports a dated retrieval snapshot and proposes a specification for evaluating a court-data slice.

Official starting points: read the source category first

The Supreme People’s Procuratorate’s January 22, 2026 typical-case release presents six personal information protection matters. Its subjects include parking applications, residential facial recognition, recruitment, doxxing, information relating to deceased people and relatives, and tourism reservations. Several accounts describe supervisory recommendations and rectification. The Hangzhou doxxing account describes a civil public interest suit filed on April 11, 2025 and a judgment on August 28, 2025.

The March 30, 2023 official release also includes civil, administrative supervisory and criminal-with-civil matters. These releases are selected explanatory accounts, not a register of every proceeding. Their publication dates should be stored separately from dates inside the narratives.

The engineering implication is straightforward: preserve the original source category before deriving a product label. An official account can be useful even when it contains no court case number or complete judgment. It should remain identifiable as an account. A search result headed “administrative public interest litigation” should not become a row labeled “administrative judgment” without evidence that the record is that document. The 2026 release illustrates why the distinction matters.

Build separate source tracks

Use these as evaluation categories, not as a claim that every source is available from one supplier. Ask vendors to state which tracks their delivery contains. A court corpus should be described as court material. If a pilot also uses official prosecutorial summaries, keep those in a separate source collection with their own acquisition and refresh rules.

The same matter may produce several artifacts. Keep a document identifier for each artifact and a separate, reviewable matter identifier when a relationship is supported. Collapsing the two makes a multi-document proceeding look like several independent disputes; merging by a common party name can create the opposite error.

What live court-record searches returned

On October 10, 2026 we queried SinoVerdict’s judgment-record index. For each phrase, we took the union of ik_smart and ik_max_word phrase searches. All count responses completed without shard failures. The final row below intersects two union queries against the same document. No document-type, year or procedural filter is applied.

The 776 records are a co-occurrence candidate pool. They are not 776 verified personal information public interest cases. One phrase could occur in a quotation and the other in an unrelated passage. Multiple judicial documents may concern one dispute. Some records may lack usable text beyond the retrieved fragment. Review is required before applying a substantive label.

The larger pools expose a practical choice. A title search is a narrower discovery route than a body search, but it can omit documents whose titles use different language. A body intersection may locate less obvious candidates while introducing incidental mentions. These observations concern query behavior; they do not establish either query’s precision or recall.

The downloadable query definitions and aggregate responses preserve the observation date and exact filters. They include no individual case text. Do not add the rows: the pools overlap, and the intersection is drawn from two of them. A future run may differ when the underlying index changes.

Do not equate a keyword with a procedural role

Suppose a retrieved record contains both target phrases. Before inclusion, a reviewer should identify who brought the claim, what interest the claim seeks to protect, which procedural track is described and whether the document resolves that claim. A mention in a cited rule is a retrieval hit; it is not evidence that the current dispute belongs to the topic.

Keep three kinds of text separate: a party’s allegation, a prosecutor’s request and the court’s finding or disposition. A summarizer that changes “requested deletion” into “the court ordered deletion” invents an outcome even when its translation is fluent. Require the output to cite the available passage that supports each role and procedural statement.

Do not infer missing evidence from the query count. The intersection was not manually audited, and this article reports no relevance percentage. If the pilot needs a validated topic dataset, define the inclusion rules and then measure the candidates against them. Our Chinese phrase-query audit explains why retrieval counts also depend on analyzers.

Use an evidence-based label design

A useful first labeling pass is deliberately modest: source type, document type as observed, procedural track supported by the text, privacy subject if identifiable, and available remedy evidence. Each derived label needs a supporting passage or an explicit unknown state. An attractive taxonomy is not a substitute for readable evidence.

Allow multiple privacy subjects and multiple remedies. Force neither into a single category merely to simplify a dashboard. Keep unknown distinct from absent: no readable passage about an injunction does not establish that no injunction was sought or granted. Likewise, a missing amount field is not a zero monetary award.

Store review status independently of the substantive label. Suggested, confirmed, rejected and unresolved records serve different purposes. A product can show a suggested match while making its evidentiary status visible. This lets a reviewer repair one field without silently changing all downstream answers that depend on the record.

Date fields need their own evidence

A source can carry a publication date and describe earlier investigation, filing, hearing and disposition dates. Those are separate events. Store the original date string, normalized value, event type and supporting span. If a narrative provides only a year or says “the same month,” retain the uncertainty instead of manufacturing a complete calendar date.

For a benchmark about a legal change, specify which date determines the cohort. A publication-year split answers when material became available in that source. A filing-date split answers a different question. A decision-date split answers another. The filing-cohort and decision-date study shows why an undifferentiated date field can distort a research design.

Freeze the source snapshot used for evaluation. If labels are updated after a new official account is published, retain both the earlier evidence state and the revised state. Otherwise a repeated test may improve because the source changed, while the team attributes the difference to its model.

Link documents without inventing a case chain

Use original identifiers when available: court case number, official account URL, document title, publisher and relevant dates. Treat a proposed relationship between records as an edge with evidence, not as an unqualified merge. The link should state what it means: the account describes the judgment, a document cites another record, or a reviewer considers them part of the same matter.

A shared party name is a discovery clue, not a confirmed relationship. Masked individual names should stay masked. Do not reconstruct identities from partial names, institutions or geographical clues. A false entity match can contaminate both research results and a training corpus long after the original join rule is forgotten.

Ask the supplier to explain duplicate handling and document survival rules. A delivery can contain multiple versions or procedural documents without being unusable, provided the buyer can distinguish them. Reporting raw records as independent disputes conceals the very information a legal research product needs.

Require a coverage report that admits missingness

This table is a proposed acceptance specification. It does not promise universal availability in SinoVerdict or any other corpus. Measure completeness and correctness separately: a field may be filled in every row and still be consistently wrong. Our derived metadata audit explains the distinction.

A pilot should test more than successful retrieval

Start with a defined frame: candidate queries, source collections, time bounds and document categories. Select review strata before seeing favorable examples. Record the randomization method, seed, sample size and duplicate policy if you sample. This article contains no sampling experiment, so the counts above support no claim about semantic purity.

Include negative and unresolved examples. Ask whether an official account actually reports a filed court case; whether a judgment resolves the public interest component; whether an amount is sought or awarded; and whether a requested deletion was withdrawn, ordered or not addressed in the available passage. The expected answer can be “insufficient evidence.”

Retain each failed question, retrieved record identifiers, available Chinese evidence, English answer and reviewer correction. Group failures by coverage, retrieval, labeling, translation and unsupported inference. A single success score hides whether the next fix belongs in the source pipeline or the model prompt.

Rates a court corpus cannot establish

Do not derive a national rectification rate, prosecution-to-filing rate or public interest litigation success rate from these record counts. Those measures need a defined population and records of events that may never reach court. A selected official release supplies examples; a captured court collection supplies another selected observation surface. Combining them does not automatically produce a valid denominator.

Absence has several possible explanations: different wording, unavailable text, a record outside the captured source, or an event that produced no court document. State the observed retrieval boundary. Do not turn a missing match into an assertion that an organization has never faced a privacy-related proceeding.

SinoVerdict’s 160M+ record description refers to its broader court-record corpus. It does not describe 160 million privacy cases, complete readable judgments or universal original-source URLs. Available court case numbers can support independent checking; they should not be presented as guaranteed working links to official originals.

From source discovery to an enterprise evaluation

Use official releases to understand the vocabulary and document tracks, then test a court slice against a buyer-defined task. Keep the source boundaries visible in the interface and contract. A well-specified pilot can discover that the intended product needs an additional source collection; that is a useful result, not a reason to relabel missing material as covered.

SinoVerdict offers enterprise court-data access through bulk, REST API and MCP workflows. Request a trial API key and scoped corpus coverage report, and review the delivery options. Specify the public interest track, years, regions, evidence fields and intended use. Confirm actual slice availability, English output and permitted uses before agreeing a production delivery.

Frequently asked questions