Regulatory data sources

China Government Procurement Complaint Data: MOF Notices and Court Records

China government procurement complaint data starts with finance-authority publications. A court judgment is a separate research object. For a legal AI product, preserve the original complaint disposition, any separately published penalty, the procurement project and subsequent judicial proceedings as linked but independent records. Combining them into a single “procurement violation” row loses the distinction your users need.

This guide identifies official starting points and proposes a collection and acceptance design for overseas data teams. It does not offer a national complaint count, a supplier blacklist or a legal opinion on any participant. The source check is dated October 18, 2026; individual examples illustrate document structure rather than representative outcomes.

1. Begin with the question your dataset must answer

A procurement lawyer may want the reasoning behind a challenged technical specification. A supplier intelligence team may want an authoritative restriction and its effective period. A legal research model may need the distinction between an allegation and the authority’s findings. These tasks can involve the same project while requiring different evidence and different labels.

Write the intended output before collecting records. “Find published complaint dispositions about qualification criteria” is a bounded research request. “Determine every supplier’s eligibility to bid today” requires a substantially different product, with current official evidence, documented scope and a process for subsequent changes. A historical court-data sample alone cannot establish that second answer.

Our related administrative penalty source guide explains the wider distinction. For procurement, the extra complication is that one project can generate multiple notices, multiple disputed issues and multiple institutional responses. Decide which of those objects is your unit of analysis.

2. Where to look for official publications

The China Government Procurement Network is an official starting point. Its information announcement page separates procurement announcements from the Ministry of Finance’s government procurement information announcements. Use the relevant collection and the issuing authority recorded on each item. A ministry series should not silently become the label “all China procurement complaints.”

The Ministry’s Order No. 94, Measures for Government Procurement Queries and Complaints, provides useful source context. Article 33 describes the contents of a complaint decision, including findings, grounds, disposition, review or litigation information and the decision date. Article 34 addresses publication of complaint handling results on designated procurement information media. These provisions help explain why a published result notice and the complete decision must remain distinct in a dataset.

Read the applicable instruments with counsel for a real matter. Here the purpose is document classification: record what you actually retrieved, rather than promoting a short public notice into a complete administrative case file. An accessible announcement does not prove that its underlying submissions, evidence or full reasoning are all publicly available.

3. A real example: notice No. 3541

Ministry of Finance information announcement No. 3541, dated September 14, 2026, is headed as a complaint handling result announcement. It identifies project CG2025-PL-GK-HW-019, concerning a Customs Administration network-security product procurement. The complaint concerns certain substantive tender requirements and alleged differential or discriminatory treatment. The published disposition states that some complaint matters were established and others lacked factual support.

That compact publication demonstrates why a binary “complainant won” field is inadequate. Preserve the distinction between the asserted issue and the disposition, and retain the mixed outcome. Do not infer a fine, supplier exclusion, compensation or a subsequent judicial result from this notice. The source check did not establish those events.

A useful extraction exercise is to ask an annotator to identify the project, source document type, allegation, cited grounds and outcome without adding an unstated consequence. Run the same exercise through your model. A plausible narrative that invents a sanction should fail acceptance even if its entity extraction is accurate.

4. Keep an issue table beside the document table

A practical data model uses a document table for retrieved publications and a separate issue table for complaint matters. Each issue links back to the publication and carries its own source-supported classification. Allow “mixed,” “not stated” and “needs review.” Do not force every record into one positive or negative outcome simply because the downstream dashboard expects a percentage.

Maintain the original Chinese wording beside normalized English labels. An English label should explain its scope, not replace the source. Record whether the wording is an allegation, a response, a finding or a disposition. This is especially important when a model reads a passage describing a complainant’s position and mistakes that position for the authority’s conclusion.

Store an evidence span or paragraph locator with each extracted claim, together with parser and translation versions. If the text changes, the derived fields should be traceable to the exact version that produced them. An audit should be able to distinguish a source correction from a parser improvement without relying on a single mutable spreadsheet cell.

5. Match identifiers without merging unlike things

Use different fields for the procurement project number, lot number, publication series number, administrative instrument number and court case number. They identify different objects. Retain the original form as well as a normalized value: spacing, punctuation and full-width characters can affect matching, but normalization must not erase substantive distinctions.

A project identifier can connect tender, result and complaint publications. It does not prove that two documents concern the same disputed issue, supplier or stage. A company name match is weaker still. Prefer explicit cross-references, then review combinations of project identity, institutional roles and dates. Save the evidence used for every confirmed link.

Represent uncertain relationships as candidate links with a reason and review status. Do not let candidate edges flow automatically into a supplier-risk score. A useful negative test is a pair of notices sharing a purchaser or procurement agent but concerning different projects: the system should keep them separate.

6. Evaluate the judicial track independently

For a court-data pilot, begin with Chinese procurement terms and known case identifiers where available. Possible exploratory terms include 政府采购, 投诉处理决定 and 财政局, combined according to the search engine’s documented behavior. These are candidate retrieval terms, not a validated filter, and no count or precision claim for them is made here.

A document may mention procurement as background, quote an earlier decision or concern a different dispute involving the same organization. Have reviewers establish what the court was deciding and what procedural stage the document represents. Preserve later proceedings as separate documents rather than treating every matching record as a new procurement complaint.

Assess missing or truncated text explicitly. A record with a court case number but insufficient substantive text may be useful for discovery while failing a reasoning-analysis requirement. The full-text coverage guide explains why field existence alone is insufficient. Require readable evidence for the particular passages your product promises to analyze.

7. Dates need a role, not just a column

Separate project year, decision date, publication date, retrieval time and the date of any later status check. The example’s project identifier includes 2025 while the notice was published in 2026. Classifying the entire record into one year without declaring which clock you use makes historical comparisons difficult to interpret.

For a publication-monitoring service, publication date can be the reporting clock. For a study of decision behavior, decision date may be more appropriate when actually available. A missing date remains missing; a URL path can be a discovery clue but should not silently substitute for the document’s stated date.

Preserve the first observed version and subsequent observations where your collection rights permit. Distinguish new publications, corrections, removals and access failures. A page becoming unavailable should create a documented observation, not an automatic declaration that the decision was revoked or the underlying conduct disappeared.

8. A pilot that can fail for the right reasons

Start with a deliberately bounded sample: named authorities, a stated period, specified document classes and known examples. Include a mixed-outcome complaint, an ordinary award notice with no complaint, an unrelated project sharing an agent, and a judicial record that merely quotes administrative material. These examples test whether the system preserves the distinctions that matter.

Measure retrieval and extraction separately. A useful document that the system never retrieves is a search failure; a retrieved document with an invented disposition is an extraction failure. Report the denominator, labeling method and reviewed errors for each. Small convenience samples cannot support a national complaint-success rate.

Agree in advance what would stop the pilot: unreliable source classification, missing evidence spans, unreviewed entity joins or unavailable substantive text. An honest failure at this stage is cheaper than shipping a confident supplier-status answer that the underlying records cannot support. For the wider handover process, use our court-data delivery acceptance checks.

9. Define commercial scope and language separately

Public accessibility and intended commercial use are separate questions. Record source terms and obtain appropriate review for storage, display, redistribution, model training, translation and downstream delivery. Do not infer a training or redistribution grant from a successful download. Administrative publications and a licensed judicial collection can require different permissions and acceptance documents.

SinoVerdict’s 160M+ headline describes court-data records; it is not a government procurement register size or a promise of complete procurement coverage. Request a scoped trial API key and corpus coverage report for the judicial component. Bulk, REST and MCP delivery requirements should be discussed against the actual product and agreed scope through our delivery overview.

Validate language behavior on the endpoint you will use. Chinese source text, an English interface, English query support, generated English summaries and a fully translated corpus are separate capabilities. The English-access capability matrix and licensing guide set out those boundaries. A procurement pilot should test its own known Chinese and English queries rather than borrowing a general marketing claim.

Frequently asked questions