Home → Insights → Patent invalidation data

China Patent Invalidation Data: CNIPA Decisions and Court Records

China patent invalidation data is spread across administrative decisions, examination services and court publications. These sources answer different questions. A published invalidation decision explains an administrative proceeding; a judgment explains judicial review; a patent-status check addresses the state visible at a particular time. An AI product needs to keep those objects separate before combining them into an answer.

This guide maps verified official entry points and a measured judicial retrieval pool. It is for legal-AI data teams and China-practice researchers evaluating coverage, rather than a guide to filing an invalidation request. Source descriptions were checked on October 14, 2026. The suggested data model and acceptance checks below are our engineering recommendations, not official requirements or legal advice.

1. Where to start: the official source map

CNIPA's February 3, 2026 answer about recent invalidation proceedings directs parties and the public to its patent examination information service for case status and oral-hearing information. The named entry point is cpquery.cponline.cnipa.gov.cn/chinesepatent/index. This verifies an official route and its stated purpose. We did not sign in or establish that every record or field is available to every user.

A separate CNIPA answer about decision reasons and results describes published invalidation decisions and the decision-search function of the Reexamination and Invalidation Department. It also distinguishes published material from documents accessed through copying procedures. Treat the answer as source guidance: it does not establish permission for automated bulk collection, unrestricted redistribution or model training.

Begin with the object your product promises to answer about. If the user asks for the latest status of a named patent, the acceptance question is whether an appropriate current official observation can be obtained. If the user asks how courts treated challenges to invalidation decisions, the relevant material is judicial reasoning. Those questions may share identifiers without sharing a source or a completeness standard.

2. A published court example shows why one patent is not one proceeding

The Supreme People's Court Intellectual Property Court published an analysis of case (2023) 最高法知行终1255号 on August 7, 2025. Its narrative describes multiple invalidation proceedings and litigation involving the same utility-model patent. It also discusses a later administrative decision following a different appeal. This is a concrete example of intertwined events, not evidence that the same pattern occurs at a particular national frequency.

For a data team, the lesson is to preserve the patent, administrative proceeding and court document as distinct objects. The analysis includes a masked patent number. Keep that masking: do not reconstruct missing digits from a title or a party-name resemblance. Store the verified court citation and the official page address for this particular published example. That does not imply that other corpus records have working original-source links.

A second official page, case (2025) 最高法知行终794号, illustrates the difference between a case's citation year, administrative-event dates and publication date. The page is dated September 4, 2026 and contains an administrative judgment describing an earlier invalidation request and decision. These are separate clocks. A dataset filtered by page publication year will answer a different question from one filtered by the administrative request date.

Neither example is a representative sample. They establish that the data model must accommodate these relationships and dates. They do not establish an invalidation rate, an appeal success rate or coverage of all published decisions.

3. What we actually retrieved from the court corpus

On October 14, 2026, we ran three Elasticsearch count queries against the judgments index. Each query uses the union of analyzed-phrase searches with ik_smart and ik_max_word, excludes dataset bench and checks that failed shards equal zero. The results and exact requests are available in the downloadable query evidence.

The pools overlap and must not be added. These are retrieved record counts, not distinct patents, administrative decisions or cases. We did not calculate their intersection, deduplicate them or perform a semantic relevance audit in this measurement. The difference between title and body retrieval therefore cannot be interpreted as the number of records a title search misses from a known relevant population.

Analyzed phrase matching is not a guarantee that every returned body contains an identical literal character string. Different ingestion batches and query analyzers can affect results. Our phrase-query analyzer audit explains why a default query returning zero is insufficient evidence of absence. A production pilot should test both recall candidates and false positives for the actual terms it uses.

These counts are useful for scoping a review queue. They do not tell a buyer how many full decisions can be delivered. Some records may lack usable reasoning; some may repeat the same document; some may concern infringement litigation that only refers to an invalidation proceeding. A delivery claim needs document-level inspection under the buyer's agreed scope.

4. How to distinguish judicial review from incidental mentions

Start by retaining the original title, court, case number and available date fields. Add a separate classification field for the question being asked: review of an invalidation decision, review of an examination rejection, infringement litigation with an invalidation reference, or unresolved. Do not silently label every record containing the authority's name as a patent invalidation case.

Review the dispute description and the court's own disposition. A body may quote the applicant's allegation, summarize an earlier administrative decision or discuss a related case. Those passages should not be assigned to the deciding court as its conclusion. For retrieval-augmented answers, return enough context for the reader to distinguish an argument, a historical event and the holding.

Use the administrative litigation dataset guide for the broader judicial scope and the IP enforcement guide for infringement-oriented research. This page focuses on source boundaries and invalidation-related records. Combining all IP disputes into one total would conceal the exact gap a patent-data buyer needs to assess.

Leave uncertain records in an explicit review bucket. A missing procedural label is a quality issue to quantify, not an invitation to infer one from a keyword. The product can still return an unresolved candidate if it explains what is known and what must be checked before relying on it.

5. Build an event graph with evidence for each edge

A useful minimum model has a patent identifier, an administrative proceeding identifier, a decision identifier and a court-document identifier. Preserve their raw forms alongside normalized values. Patent application, publication and grant identifiers should have typed fields; a string that resembles another identifier should not automatically be substituted for it.

Attach dates to events rather than to one generic year column. Where available, distinguish request, administrative decision, court filing, court decision, official publication and collection dates. Mark extracted dates as derived and retain the source passage. Unknown dates remain null. Sorting by collection time may help a pipeline process updates, but it does not turn the newest collected document into the newest legal event.

Every relationship should identify its support: an explicit patent number, a cited administrative decision number or a stated court case reference. A party-name or title match can create a candidate edge for review. It should not become a confirmed link without documentary corroboration. Masked identifiers remain masked and unresolved links remain visible.

Multiple court documents can relate to the same administrative decision, and multiple decisions can concern the same patent. Deduplicating by patent number would erase the history. Deduplicating only by court case number can also remove distinct documents. Keep a record identifier, a document-level duplicate key and a proceeding-level grouping key with separately documented rules.

6. Acceptance checks for an invalidation-data pilot

Write the acceptance statement before requesting a sample. Specify the judicial question, courts or jurisdictions, time fields and document classes. A pilot for administrative-review reasoning should not pass merely because it returns many infringement judgments containing an invalidation reference.

1. Source separation: identify which items are administrative decisions, court documents, summaries or status observations. Each item needs a source type rather than one generic judgment label.
2. Scope reconciliation: report retrieved records, reviewed relevant documents and grouped proceedings separately. Explain exclusions and unresolved records without using a single large total as a substitute.
3. Text usability: inspect whether the reasoning and disposition are present. An existing body field is insufficient; disclose missing, placeholder and truncated text under the agreed measurement.
4. Relationship accuracy: test patent-to-decision and decision-to-court links with evidence. Record ambiguous identifiers and unmatched references rather than discarding them.
5. Temporal meaning: reconcile the chosen date field against source passages. State whether the window describes filing, decision or publication events.
6. Reproducibility: deliver query definitions, snapshot dates, filter versions, extraction versions and review criteria with the pilot report.

If the pilot uses a sample, record its population, random seed, sample size and stratification. A convenience sample of prominent cases cannot estimate the whole corpus's relevance or missing-text rate. Include ingestion-batch and time coverage, and disclose dominant parties or litigation clusters. Any subgroup with fewer than thirty reviewed examples is insufficient for a statistical conclusion here.

For detailed delivery mechanics, use our court-data acceptance checklist. A count query is an initial scope check. A successful pilot requires evidence that the delivered records can support the specific answers the customer intends to publish.

7. What an AI answer should expose

Suppose a researcher asks how courts handle multiple invalidation challenges involving one patent. A useful answer cites the verified judicial document, distinguishes the referenced administrative events and states the limits of the retrieved evidence. It should not label a prior decision as the current status just because that decision appears last in the retrieved excerpt.

Expose the Chinese case-number citation and court alongside an English summary where supplied. Mark the summary as a derived aid and keep the original-language evidence accessible within the agreed delivery. A translated description of a claim or outcome needs review before it is used in a consequential patent assessment.

For current status, give the observation's source and time or explicitly state that no current official check was performed. Do not turn a search with no relevant results into a clean-bill finding. Query limitations, missing identifiers and unpublished material are different reasons why a search may be inconclusive.

8. Where SinoVerdict fits

SinoVerdict offers scoped evaluation of Chinese court records through bulk delivery, REST API and MCP access. Its 160M+ headline is a corpus record count, not a count of distinct patent disputes or a guarantee of complete text. The three measured pools above establish candidate judicial retrieval capacity only.

Ask for a trial API key and a corpus coverage report specifying invalidation-related judicial records, date semantics, necessary text and identifier checks. Treat access rights, permitted uses, redistribution and update obligations as separate written procurement items. Public availability of a source does not establish every proposed downstream permission.

A court-data partner can help organize judicial evidence. Current patent status and administrative-decision completeness still need their own source-specific verification. Keeping those acceptance items separate makes the pilot reviewable: each promise has an object, a date and evidence that can be checked.

Frequently asked questions