The Case Number Is the Only Honest Province Field

Yesterday we published a provincial axis for walking a capped Chinese court registry and ended the piece with an admission: we had built our thirty-two-character alphabet out of abbreviations a reasonable person would guess, we had found one of them wrong by accident, and we had not audited the rest against any authoritative table. Today we audited it — not against a table, against 19,591 case numbers written by the courts themselves, drawn from our own corpus.

The alphabet survived. What did not survive is the field sitting next to it. In the same pass we measured our own province labels and found that 22.3 per cent of records across all thirty-one provincial units are unreachable by that province’s official name, that roughly 11 per cent fall into no region bucket at all, that our own facet list advertises region names which return zero results when you filter on them, and that a handful of judgments are filed under provinces they have nothing to do with — including one whose court name literally contains the correct province.

All of that is our data. We are publishing it because the interesting part is not our defect list. It is that the case number catches every one of these errors and the province field catches none of them, and that is a property of Chinese court data generally, not of us.

What a Chinese case number actually is

From 2016 onward — the year our own data shows the change land, below — a mainland Chinese case number has a fixed shape: a year in brackets, a court code, a case-type abbreviation, a sequence number. (2025)内0403民初5631号 is the 5,631st first-instance civil case of 2025 at court 0403 in the region whose code is 内 — Inner Mongolia.

Two things follow. The first is that the number is self-certifying. It was assigned by the issuing court and travels inside the document text; it is not a label some pipeline attached afterwards. The second is that the first character after the year is a region code, so if a record’s province field and its case number disagree, one of them is a derived guess and you can tell which.

That is the whole audit. It costs nothing, it needs no reference table, and it works on any sample of any vendor’s output, including ours.

Finding 1: the alphabet is thirty-two characters, and only since 2016

We sampled oldest-first across all thirty-one provincial units, deduplicated by case number, and bucketed 19,591 numbers by the year inside the brackets. For each year we counted how many distinct characters appear in the first position after the year, and what share of them equal the region code of the province we had filtered on.

The alphabet collapses from 435 distinct characters to 38 in a single year, and the match rate goes from 7.1 per cent to 98.2 per cent across the same boundary. That is the 2016 case-numbering reform, measured rather than cited.

Before it, courts numbered cases with their own abbreviations, not their province’s. In our oldest-first stratum the most common first characters include 沈 for Shenyang, 昆 for Kunming, 哈 for Harbin, 拉 for Lhasa, 并 for Taiyuan, 筑 for Guiyang — city codes, prefecture codes, and in a few cases district codes. Across the 6,162 parseable case numbers in our oldest-first stratum we counted 388 distinct first characters, and we have no reason to think that set is closed.

So the honest correction to yesterday’s piece is not that the alphabet was wrong. It is that the alphabet has a start date. A provincial axis built on the first character is a sound partition key for 2016 onward and is close to meaningless before it — and a corpus that spans both eras will let you build the axis, run it, and get a clean-looking result that quietly ignores everything older than 2016.

Two characters that are not provinces

Two codes in the modern alphabet do not correspond to any province. 最 belongs to the Supreme People’s Court — (2024)最高法执监648号 — and in our index those records carry a province field whose value is the court’s own name rather than a province. 兵 belongs to the courts of the Xinjiang Production and Construction Corps, which sit inside Xinjiang geographically and number their cases under their own code: in our last-three-years sample for Xinjiang, 198 of 200 case numbers began with 兵 and two with 新. One province, two characters. Any grid that assumes one code per region loses whichever one it did not think of.

Specialised courts do not add to the alphabet. Maritime, railway transport and financial courts keep the province code and take a distinctive court number instead — (2024)辽72执恢129号 for Dalian Maritime Court, (2025)鲁7101执359号 for Jinan Railway Transport Court, (2025)京74民辖终118号 for the Beijing Financial Court. Useful to know before you conclude your alphabet is leaking.

Finding 2: the official province name reaches only part of a province

Our index stores the province in more than one form. Some records say 北京, some say 北京市. Some say 广西壮族, some say 广西壮族自治区. Filtering on the official full name returns only the records stored in that form.

We ran the same broad query for every provincial unit twice, once on the short stem and once on the official administrative name, and took the exact result totals:

Every one of the thirty-one loses at least 14.1 per cent. The aggregate gap is 24,493,889 records. Nothing about the query says so: both spellings return HTTP 200, both return real judgments from the right province, and the narrower one simply returns fewer.

You can reproduce the shape of this on our public key-free endpoint without a key, using a slice small enough to escape its ten-thousand-hit counting cap:

curl -sG https://tob.wenshucha.com/api/cases/search \
  --data-urlencode 'q=工伤' --data-urlencode 'province=西藏'
# => "total": 398

curl -sG https://tob.wenshucha.com/api/cases/search \
  --data-urlencode 'q=工伤' --data-urlencode 'province=西藏自治区'
# => "total": 324

Same corpus, same query, one word longer in the filter, 74 records fewer — 18.6 per cent of Tibet’s work-injury cases, invisible to anyone who typed the province’s legal name.

Finding 3: our facet advertises values that filter to nothing

The region facet behind our search returns thirty-four keys. Three of them are not administrative names of anything: 广西壮族省, 新疆维吾尔省, 宁夏回族省. On one query a fourth appeared — a court’s full name with the character for “province” stuck on the end. The pattern is a display layer appending a suffix to a stored stem and having no case for autonomous regions.

The consequence is worse than cosmetic. Filtering on those advertised labels returns zero. The facet offers a value; the filter has never heard of it. On a divorce query the malformed Guangxi key carried 86,880 documents by the facet’s own count and 0 by the filter’s.

And the facet is capped at thirty-four entries — we tried five different size parameters and none moved it. Thirty-four is close enough to the number of provincial units to look complete. It is not: on a work-injury query, the four malformed keys occupied slots and pushed Tibet off the list entirely, while a direct filter shows Tibet has 398 such records. A buyer scanning that facet would conclude, reasonably and wrongly, that we hold no work-injury cases from Tibet.

Finding 4: about eleven per cent of records land in no region bucket

Summing every bucket the facet returns and subtracting from the query total gives the share of records that fall into no region bucket:

Part of that is the facet cap eating the tail, and we say so rather than attributing all of it to missing metadata. But the empty province is real and we have seen it directly: Wuhan Railway Transport Court records in our index carry an empty province field while their case numbers begin with 鄂. The case number knows; the field does not.

Finding 5: some province labels are simply wrong

Filtering on one province and reading the case numbers turns up records whose code belongs to a different one. Four we can name:

The third one is the instructive case: the court’s own name begins with the characters for Hunan Province and the record is filed under Fujian. The fourth suggests the mechanism — 广德, a city in Anhui, matched against 广东, a province — a fuzzy string match on the court name, which is exactly how a province field gets derived when it was never supplied.

Rate, honestly bounded: across 6,180 case numbers in a last-three-years stratum, 252 (4.08 per cent) had a first character other than the filtered province’s code. 198 of those are Xinjiang’s legitimate 兵 and not errors at all. The remainder is 54 records, 0.87 per cent. That is a page-walk in relevance order, not a random sample, so treat it as a demonstration that the class exists and that the case number detects it — not as our corpus-wide error rate, which we have not measured and are not going to imply.

The same record has two provinces depending on which door you use

Our public case endpoint reports its own repair counters per page, and one of them is a count of rows whose province was backfilled at read time. Across six broad queries it fired on 5 of 116 rows. So the endpoint derives a province from the court name when the stored one is missing, while the raw index that powers the facet and the filter does not.

That is the deepest version of this problem and it is not specific to us. A region label on a Chinese judgment is almost never a stored fact; it is a derivation. Different doors into the same corpus derive it at different times with different rules, so “how many Guangxi cases do you have” has several true answers and the difference between them is not documented anywhere by default. The case number is the one field in the record that no pipeline invented.

Finding 6: the brackets are not normalised, and that is a silent filter

Case numbers in our corpus use both the ASCII bracket ( and the full-width bracket （. The mix moves with era but is nowhere clean:

Per province the recent mix is far less tidy than the aggregate suggests. In the same last-three-years stratum, Guangdong came back 109 ASCII to 91 full-width and Fujian 120 to 80. A parser anchored on ^\( would drop about 45 per cent of that Guangdong sample and essentially the entire pre-reform archive, and it would drop them the way these things always drop: no exception, no warning, a smaller number at the end of the run.

We also found case numbers that are empty strings, case numbers that begin with neither bracket, and one where a masked character stands in the position of the court code. If you are keying anything on the case number — deduplication, joins, citation resolution — normalise the bracket and the digits before you compare, and count what fails to parse instead of skipping it.

The audit, in five steps, on anybody’s corpus

1. Filter by each region twice — once by the short name, once by the full official name. If the totals differ, the field holds more than one spelling and one of your filters is silently narrower.
2. Filter on every value the facet offers. Any facet value returning zero is a rendered label, not an index value. Count the facet’s keys; if the count equals the number of regions you expect, ask what the cap truncated.
3. Sum the facet buckets against the reported total. The gap is records with no region, plus whatever the cap hid. Both matter and they are different problems.
4. Read the first character of the case number and compare it to the region label. Expect a real mismatch for Xinjiang Corps courts and Supreme People’s Court records; anything else is a label error, and the case number is the side that is right.
5. Bucket by the year in the brackets. If the vendor claims a provincial partition, ask what it does before 2016, when the first character was a court abbreviation out of a set of several hundred rather than a region code out of thirty-two.

Every step is a handful of queries against a search interface. None of it requires the vendor’s cooperation, a sample export, or a reference table. Run it on us: the two curl commands above are the whole method in miniature, and the Guangxi facet value that returns zero is still returning zero as we publish.

What we are not claiming

• We did not measure another vendor’s corpus. Nothing here says our metadata is better or worse than anyone else’s, and the defects listed are ours.
• The 22.3 per cent and the bucket sums are exact result-set totals for the queries stated. They are not statements about the whole corpus, which we did not query in full.
• The per-year table, the bracket table and the mismatch rate come from bounded page walks in the order the index returned. Ordered samples, not random ones. The 2019 and 2023 buckets were too thin to report and are omitted.
• The 0.87 per cent label-error figure is a floor from one ordered sample. We have not measured a corpus-wide error rate and are not implying one.
• “The case number is authoritative” means the issuing court wrote it. It does not mean it is free of transcription damage — we found masked and unparseable ones in our own data.
• We have not repaired any of this yet. This page is a measurement, not a release note; when the fixes land they will appear in our change log with dates.

The reason to publish a defect list rather than a capability list is that everything above is checkable in about ten minutes, and a vendor who tells you where their metadata bends is cheaper to evaluate than one who makes you find out during a pilot.