Short answer: copying rose from 2013 to 2018, held steady to 2023, and is lower in the 2025–26 restatements. That last step rests on 30 documents, so it is a lead to watch, not a finding. This is exploratory work outside the pre-analysis plan (PAP 07ad5da), which governs only the responsiveness claim.
Why it matters
Luo, Wang and Yang ("Laboratories of Autocracy", NBER w34219) find that after 2013 local governments became about twice as likely to copy central directives word for word, on data that end in 2020. Our corpus runs to October 2026 and holds each province's restatement next to the directive it names. So it can show whether the copying continued after the 2024 Fair Competition Review Regulation and the 15th-plan documents, a period no academic dataset covers.
Method
- Pairs: the 1,381 basis responses behind the Insights transmission figures. Each is a provincial-government document whose opening names a State Council directive as its basis, from the six provinces whose provincial-government documents are listed in full.
- Exclusions: 26 pairs where a text is under 500 characters or not held, leaving 1,355.
- Measure: the share of the response's 12-character sequences (whitespace removed, text plus attachments) that also occur in the directive. A response copied whole scores 1.
- Forwarding notices (转发) reprint the directive and score near 1 (median 0.68–0.79), so they are reported apart (165 pairs). The figures below are for the other 1,190, the provinces' own documents.
- Code:
manage.py copying_scanwritesreports/data/copying.csv(one row per pair) andreports/copying.json.analysis/copying_models.pyfits the models and writesreports/copying_models.json.
Results
| Response year | Restatements | Median share copied | 95% interval (bootstrap) |
|---|---|---|---|
| 2013–15 | 257 | 13.9% | 11.1–15.6% |
| 2016–18 | 641 | 16.8% | 15.1–18.3% |
| 2019–21 | 174 | 15.3% | 13.4–17.6% |
| 2022–23 | 67 | 14.9% | 10.4–19.2% |
| 2024–26 | 51 | 7.9% | 4.6–13.6% |
By single year after 2023: 2024 12.2% (21 documents), 2025 8.9% (17), 2026 3.4% (13, to October).
Models. Median regression on the share, controlling for response lag, the lengths of both texts and the province:
| Cut-off (responses from this year on) | Responses after | Change in median share | 95% interval |
|---|---|---|---|
| 2022 | 118 | −1.3 points | −4.8 to +2.2 |
| 2023 | 63 | −5.3 | −9.9 to −0.8 |
| 2024 | 51 | −6.5 | −11.4 to −1.5 |
| 2025 | 30 | −8.0 | −14.5 to −1.5 |
| 2024, leaving out 2026 | 38 | −3.8 | −9.7 to +2.0 |
An OLS on the log share, with a linear year trend and standard errors clustered by directive, gives the same sign for the post-2024 term. The trend itself is slightly upward.
Reading. The rise to 2016–18 matches the direction of Luo, Wang and Yang. The recent fall comes from the 2025–26 documents: leaving out 2026, the post-2024 difference is not significant. It rests on 30 documents answering 20 directives in six provinces, and those responses are the quick ones (later responses to 2025–26 directives have not been issued yet). So it cannot yet be attributed to the 2024 Regulation or to anything else.
Caveats
- Six provinces, provincial-government documents only; departments and cities are not in the pairs.
- A basis response names the directive in its opening. A document that restates a directive without naming it is not counted.
- Copying measured this way is textual. A province can restate a directive's substance in new words, or copy its words while changing the substance. Hand-reading a sample of low-share and high-share pairs is the next step.
- 2026 is a partial year.
Next steps
Re-run copying_scan after each China-side crawl. The 2025–26 cohort will fill in as slower provinces answer. If the lower copying holds with 60 or more recent pairs and a hand-read sample, it becomes a finding worth an Insights card.