The 170M Judgments Problem: How to Structure PRC Court Data for AI

China has published the largest body of court decisions in the world — more than 170 million judgments, and still growing. For a legal AI team building China coverage, that number reads like an asset. In practice, raw volume is closer to a liability. A pile of 170 million heterogeneous documents is not a corpus; it is a sourcing problem, a cleaning problem, and a retrieval problem stacked on top of each other. The hard part of Chinese legal AI was never getting the text. It is turning the text into something a model can retrieve, filter, and cite without lying.

This piece walks through what actually stands between a raw judgment dump and an AI-ready corpus: the shape of the source material, why scale makes every cleaning decision harder, the structuring pipeline, the fields that matter, and what to demand from any vendor that claims to have done the work.

What a raw Chinese judgment actually looks like

Chinese judgments have one enormous advantage and several practical traps. The advantage: they are highly formatted. Courts draft to standardized conventions — a caption with the parties and case number, procedural history, findings of fact, reasoning, the operative holding, and the bench. That regularity is exactly why the corpus is so amenable to structured extraction; it is the single biggest reason Chinese case law is more tractable at scale than free-form text from many other jurisdictions.

The traps sit underneath that regularity:

• Metadata lives in prose, not fields. The court, the cause of action, the outcome — all of it is written into the document, not handed to you as columns. Extraction has to read the text the way a clerk would.
• The same matter appears more than once. First instance, appeal, retrial, enforcement; original and amended versions; reposts across batches. Without de-duplication and versioning, a single dispute can inflate your counts and skew any statistic you compute.
• Redaction is built in. Under Supreme People's Court publication rules, judgments keep party names but redact identifiers like ID numbers and addresses. That shapes how entities can be resolved and reminds you that this is sensitive data with a compliance frame around it.
• Format and encoding drift. Documents arrive from many courts and publication periods, with layout variation and occasional OCR or encoding noise. A rule that works on 95% of documents still fails on millions.

Why volume makes everything harder, not easier

It is tempting to treat 170 million as a marketing number. For an engineering team it is a constraint. Three things break at that scale that would be trivial at 170 thousand:

• Every rule meets its exception millions of times. A field-extraction heuristic that is 99% accurate is wrong on well over a million documents. Quality has to be measured and monitored, not assumed.
• Heterogeneity is the norm, not the edge case. Across thousands of courts and two decades of publication, there is no single template. Structuring has to be robust to drift rather than tuned to one layout.
• History cannot be re-crawled. Publication volume on the public repository peaked around 19.2 million documents in 2020 and fell sharply before partially recovering; the 2021–2023 vintage entered the record thin. A corpus assembled across the full window is a snapshot that today's feed cannot reconstruct — so the structuring has to preserve and de-duplicate that history correctly the first time. We cover that availability curve in The State of Chinese Case Law Data, 2026.

In other words: the more documents you have, the more the value depends on structure, because the cost of disorder scales with the pile.

The structuring pipeline, layer by layer

Turning raw judgments into an AI-ready corpus is a pipeline, not a single step. Each layer makes the next one possible.

None of these layers is glamorous, and that is the point. The differentiator in Chinese legal data is not who can crawl the most pages; it is who has done layers 2 through 6 correctly across the whole corpus. That is the same engineering problem we frame from the buyer's side in Building China Coverage Into Your Legal AI.

The schema: fields an AI corpus has to expose

A judgment is useful to a model only when the things a lawyer would filter on become things a query can filter on. At minimum, a structured PRC corpus should expose:

• Stable record ID + official case number (an-hao) — the anchor for citation and de-duplication.
• Court and court level — basic, intermediate, high, or the SPC; essential for weighting authority.
• Document type — judgment, ruling, mediation, or enforcement decision; they answer different questions.
• Date and procedural posture — first instance, appeal, retrial; without it you can't tell precedent from procedure.
• Cause of action (an-you) — the controlled vocabulary that lets you isolate a practice area cleanly.
• Parties, de-identified — per SPC redaction rules; enough to disambiguate, not to re-identify.
• Statutes and articles cited — the bridge between facts and law.
• Operative outcome — what the court actually ordered.
• Segmented full text — facts, reasoning, holding kept separable, so retrieval returns the right passage, not the whole document.

Expose those fields cleanly at 170M scale and the corpus stops being a search box and becomes infrastructure: you can facet by court and year, isolate a cause of action, weight by authority, and ground every answer. We go deeper on the field model and integration patterns in our Chinese case law API structure walkthrough.

Citation grounding is the whole point

For legal AI, the difference between a usable system and a dangerous one is whether every surfaced fact resolves to a real document. A retrieval-augmented system that can show case number, court, date for each passage is defensible; one that paraphrases plausible-sounding Chinese law from model weights is a liability waiting for a malpractice question.

Grounding is not a feature you bolt on at the end — it is a property of the data. It requires clean full text (layer 1), stable IDs (layer 2), and a preserved link back to the original judgment (layer 5). If those are missing, no amount of prompt engineering will make the answers traceable. This is why "we have the documents" and "we have a corpus your AI can cite from" are entirely different claims.

What to demand from a structured corpus

If a vendor offers you Chinese case law, volume is the least interesting thing they can tell you. These are the questions that separate a data pile from an AI-ready corpus:

1. Which fields are extracted, and how accurate are they? Ask for the schema and a measured accuracy figure on case number, court, cause of action, and outcome — not a claim, a number.
2. How is de-duplication and versioning handled? Same matter across instances should be linked, not double-counted.
3. Is the cause-of-action taxonomy normalized? You should be able to isolate a practice area without chasing label variants.
4. Does every record link back to the source? Grounding is non-negotiable for legal use.
5. What does the English layer cover — query, summaries, fields, or all three? Translation of a few thousand documents is not coverage.
6. How is the corpus delivered and updated? Bulk to seed, API to serve, incremental sync to keep it fresh.

The bottom line

The 170M number is real, and it matters — no other jurisdiction publishes case law at this scale. But the number is the raw material, not the product. The asset is what survives normalization, extraction, de-duplication, taxonomy, citation linking, and a cross-lingual layer, measured and maintained across the whole pile. That is the work behind a corpus an AI team can actually build on, and it is the work behind SinoVerdict: a structured, machine-readable corpus of more than 170 million Chinese judgments, with English query and summaries, structured fields, and cited links back to the original documents — delivered as a bulk dataset, a REST API, and an MCP server.