Data

Coverage over time

Documents held since 1 October 2026, by the day each was added: 208 then, 50,256 now. Provinces with 100 or more local documents: 0 then, 31 of 31 now. Growth here is collection, not policy output.

Documents held, by level

Provinces with 100+ local documents (of 31)

Corpus [csv]

How the counts on other pages relate to each other.

Policy documents by administrative level [csv]

Level of the issuing body's government (central, province, prefecture, county). Unplaced: the issuer is not yet linked to a division.

Policy documents by year [csv]

Year of the document's own date (成文日期 where found, else the portal's posting date). Collection is uneven: recent years and the crawled provinces are over-represented, so counts are not a measure of policy output.

112 documents dated before 2000 are not shown.

Local policy documents by province [csv]

Documents issued by a province or by a government inside it. Provinces whose portals were enumerated in full have far more documents than those reached only through citations or archives: compare provinces with care.

How documents were retrieved [csv]

direct crawl: fetched from the government site; commoncrawl / wayback: read from a web archive copy of the government page, without contacting the site.

Field completeness [csv]

Share of the 50,270 policy documents with each field filled.

Text length (characters): p10 313 · median 2,126 · p90 7,597 · p99 28,173. Posting lag where the own date differs: median 9 days, IQR 4–25, p90 173. 26,311 documents carry the same own and posting date.

Citations found, by method [csv]

A citation is resolved when the cited document is held. Unresolved citations are kept: ranked by how often they are cited they form the crawl frontier.

Citation-graph degree distributions [csv]

Over the 50,270 policy documents and the 24,630 resolved document pairs. Heavy tails are expected: a few State Council directives are cited hundreds of times.

Who cites whom [csv]

Resolved citations by the level of the citing and the cited document, with the time between them (citing date minus cited date). Negative lags are references to a document dated later, usually a draft or a dating error.

Most-cited held documents [csv]

Distinct held documents citing each one.

Crawl frontier: most-cited documents not held [csv]

75,166 distinct documents are cited but not held. A title-only reference may name a document we hold under a different title.

Unequal-treatment provisions, by mechanism [csv]

Provisions that treat firms differently by ownership, origin or location. "Bans" are provisions that prohibit such discrimination.

Confidence scores are not informative yet: 100.0% of provisions carry the screen's default of 0.40.

Policy measures, by instrument [csv]

Each measure is one instrument found in a document, with how binding it is.

Stated rationales [csv]

Why a measure says it exists, each backed by a quote.

Topic trees: size and depth [csv]

Depth = citation hops from a root document (0 = the root); a document reached from several roots counts once, at its shallowest depth.

Measure screen precision, by instrument [csv]

Ten randomly drawn screened measures per instrument, read against the quote and the document (reports/measure_validation_2026-10-25.csv). Checked by the AI assistant, not by an independent human coder; ten per class is indicative only.

Instrument rows were checked before the 25 October fix of the unequal-treatment screen, which removed most false local-production, local-content and own-brand hits (the unequal-treatment rows show the re-check after it). Procurement preference still matches 本地区 outside any condition.

Hand validation [csv]

Random samples coded by hand against the source text. Precision only: how often a hit is right, not how many true cases are missed.

Responsiveness models: Cox diagnostics [csv]

Time from a State Council directive to a provincial document citing it as its basis, stratified by province, standard errors clustered by directive.

Recall proxy: of 1,288 pairs a title echo finds, the citation measure also finds 1,069 (83.0%).

Known gaps

• Coverage is uneven. Six provinces' government portals were enumerated in full. Seven more were crawled from their own portals' listings, and the rest come from web-archive copies (Common Crawl): those are partial (the Provinces page marks each). Never compare raw counts across provinces.
• Some documents have no body text (see Field completeness). Many Chinese government sites refuse access from outside mainland China, so their pages are re-read in China-side sessions (816 filled on 2026-10-21); scanned attachments are read by OCR, which garbles some characters.
• Title citations are matched by normalised title. A reference to a generic title may resolve to the wrong year's version; references to documents we hold under a different title stay unresolved.
• One copy stands for each numbered document: the issuer's own if held, else the earliest. An unnumbered repost with the same title and nearly the same text is folded into it; a re-issue with its own title or text is kept (re-issuing is dissemination).
• Dates: the document's own date where the text gives one; otherwise the portal's posting date, which can lag by weeks or years (Field completeness gives the share).
• Issuers are placed by name, then by site. A body named without its place (经信委, 市场监督管理局, 自治区卫生健康委) is placed by the government site it came from; until 25 October 2026 about 1,150 such documents were counted as central. Documents still unplaced are listed as such.
• Screens are keyword- or model-based. Precision is hand-checked (Hand validation); recall is not, beyond the title-echo proxy. Screen confidence scores are not yet informative.