Licensing Chinese Court Data: Rights, Delivery and Buyer Acceptance

A Chinese court-data purchase needs two separate acceptance records: what is delivered, and what the buyer is permitted to do with it. A working REST endpoint proves neither redistribution rights nor permission to train a model. This guide gives legal AI buyers a practical checklist for evaluating SinoVerdict or another supplier without treating a product description as a grant of rights.

SinoVerdict offers enterprise discussions around a 160M+ Chinese judgment-record corpus and structured access. The headline is a record count, not a promise of distinct cases, complete bodies or full English translations. The scope of a particular export, account or license must be established separately. This page is a procurement guide, not a license agreement, quotation or legal opinion.

Put the intended use before the price

Describe the actual data flow before asking for a quote. An internal research team reading a small evaluation slice, a legal product displaying excerpts to subscribers, and a model developer training on a bulk export are different projects. The differences include who sees the material, where copies remain, which outside service receives it and what persists after the agreement ends.

Write down the buyer entity, authorized users, hosting locations, subprocessors, intended outputs and retention period. Include caches, search indexes, embeddings, logs and backups rather than discussing only the original files. This does not establish that a use is permitted. It gives both sides a concrete workflow to evaluate and describe in the agreement.

The public product and service documentation reviewed for this update establishes access surfaces and technical boundaries. It does not establish a blanket grant for the uses in this table. No executed buyer agreement was used to fill those gaps. Treat each unanswered permission as unconfirmed until the relevant agreement and rights evidence resolve it.

Check the supplier's authority for the proposed scope

Ask what the supplier is offering rights in: collected records, a processed database, extracted fields, annotations, translations or hosted access. Then request the evidence and representations relevant to that offer. A statement about transformation or engineering does not by itself resolve rights in the underlying material. A contract should identify its subject clearly enough that buyer and supplier are discussing the same artifact.

Do not infer an unrestricted license from public availability, a downloadable file, a robots rule, or a successful request. Do not infer that one supplier's authorization extends to a different dataset sold under a similar name. The appropriate review depends on the material and intended use; qualified counsel should assess the actual transaction. This guide does not certify the acquisition-rights chain of any particular corpus.

Make the delivery schedule measurable

A useful technical schedule names the snapshot, date of measurement, included collections, format and field definitions. For an API it should name the endpoint and schema version. For a bulk delivery it should identify the manifest and integrity-check method. A sample downloaded during a sales call is not a replacement for the final deliverable's manifest.

Separate raw records, distinct case numbers and deduplicated document units. One case number can cover several procedural documents, while ingestion batches can contain duplicate records. Ask the supplier to explain its grouping key and survivor rule rather than treating all repeated case numbers as errors. Document how missing or malformed case numbers affect the reconciliation.

Body availability also needs its own measurement. A nonempty field can contain an excerpt, a publication explanation or a truncated judgment. Request the count of records with usable source text and the rule used to identify them. Review examples from different years and source batches. Do not convert a schema's body_text field into a claim that every delivered record contains a complete judgment.

For structured fields, request fill rates and reviewed accuracy separately. A populated party or disposition field may be a best-effort extraction rather than an authoritative court field. Agree how uncertain extractions and missing values appear. The field census illustrates why year and source-batch differences belong in acceptance rather than a footnote.

Confirm the interface you will actually receive

The public labor demo, full-corpus Chinese retrieval and current remote MCP service are distinct surfaces. The remote service documents search_judgments, get_judgment, verify_case_number and verify_quote. An older local adapter may advertise a different tool set. Require a test on the interface and account proposed for your deployment.

The English access guide distinguishes Chinese text, English query handling, English summaries and full translation. English labels do not prove that the endpoint translates queries or returns English bodies. A buyer who needs English output should specify its generator, source linkage, review process and permitted downstream use.

Record service errors and quota responses during evaluation. Agree which failures count toward usage, how limits are communicated, and what the buyer's application should do when text is unavailable. This article does not quote a universal price, unlimited access, pilot duration or response-time commitment. Those are terms to confirm for the proposed account and scope.

Keep evidence checks within their actual scope

A corpus match shows that a record or passage was found in the supplier's stored data. It does not establish that the supplier has every document in the proceeding, that the record is the latest version, or that an English summary is legally authoritative. For high-consequence use, define the additional source checks and human review required before relying on a result.

The full-corpus interfaces do not provide usable original court URLs as a general guarantee. Keep available case numbers, courts and dates, and plan any required verification separately. Do not generate plausible-looking source links from identifiers. If a sample endpoint contains a working court link, verify its scope before extrapolating that behavior to another endpoint or the bulk corpus.

Publication, personal information and withdrawal need review

The Supreme People's Court's 2016 publication rules address nonpublication, masking, deletion of specified information and withdrawal. They do not establish that every stored copy is fully anonymized or remains publicly available. A buyer should ask how the proposed corpus represents those conditions and how corrections or removal notices are handled.

The Personal Information Protection Law, including Article 27 on publicly disclosed personal information, is relevant material for counsel's review. Public disclosure does not settle every downstream processing question. This guide makes no determination that a particular transfer, training project, publication or profiling use is permissible.

Give counsel the actual fields and sample handling plan, intended recipients, hosting locations and proposed uses. Ask the supplier for applicable representations, exclusions and operational procedures in writing. Avoid substituting “public data” or “anonymized” as an unsupported answer to all of those questions.

Specify updates, removals and the end of the agreement

Ask whether the purchase is a static snapshot, periodic replacement or incremental feed. Agree how new records, corrections, re-extractions and removals are identified. An increase in raw volume can include duplicates or backfills; it is not automatically evidence of newer judicial coverage. Measure freshness on the slice your product uses, with an explicit date field and measurement period.

Map a removal or correction through the buyer's system: original files, relational storage, index, cache and generated artifacts. Discuss what can be located and updated, what must be retained for audit, and which obligations survive termination. For models or translations, identify the intended treatment before creating the derivative rather than waiting until a deletion request arrives.

At termination, specify permitted retention, deletion verification, ongoing access and customer-facing continuity. Include the handling of evaluation results and reusable query sets. A clause covering source files alone leaves important operational questions unanswered, even if no additional raw records will be downloaded.

A buyer acceptance packet

1. Use and permission matrix: each intended use, the corresponding agreed clause, and any unanswered question.
2. Delivery manifest: snapshot or endpoint identity, measurement date, record units, formats and integrity checks.
3. Coverage report: relevant years, regions and document types; duplicate methodology; missing and truncated text.
4. Reviewed sample: relevant, irrelevant, missing-body and extraction-error examples with the evaluation method.
5. Language evaluation: paired English and Chinese queries, summary or translation samples, and responsibility for generated output.
6. Operating schedule: limits, update artifacts, correction/removal handling, retention and termination.

Keep these artifacts attached to the same proposed scope. If the supplier changes the endpoint, sample, corpus slice or intended rights, identify which tests and terms need revisiting. Passing a technical test should not silently fill a permission gap, and an agreed permission should not silently erase a failed completeness test.

To discuss a SinoVerdict evaluation, use the access request or email chenjiaxin@wenshucha.com. Include your required slice, interface, language output and intended uses. Request the technical sample and proposed terms together so the final decision rests on a defined deliverable and a reviewed agreement.