# Datasheet: Arieta Data Citation Graph

Snapshot 2026-11 (November 2026). Generated 2026-11-14T00:02:12Z.

## Motivation

**For what purpose was the dataset created?**
To give AI developers licensable, structured European legal data for training, retrieval-augmented generation and evaluation. Arieta Data Citation Graph is a typed, directed edge list linking EU legislation, CJEU case law and national law: references, amendments, repeals, corrigenda, implementation, legal bases, proposals and national transposition.

**Who created the dataset?**
Arieta Labs Ltd (England & Wales, company no. 16629785), from the corpus maintained by its LexAPI service.

**Who funded it?**
Arieta Labs Ltd. No external funding.

## Composition

**What do the instances represent?**
One record per directed, typed edge.

**How many instances are there?**
801,149 edges as of the November 2026 snapshot (snapshot id 2026-11).

**How were the counts made?**
Documents and citation edges are exact counts. Articles and tokens are estimated from a 2% random sample of rows, extrapolated per lane (small lanes are counted exactly); tokens are characters ÷ 4. Withdrawn documents and held national case law are not counted.

**Which jurisdictions?**
European Union (EU), Austria (AT), Czech Republic (CZ), Germany (DE), Denmark (DK), Spain (ES), Finland (FI), France (FR), Croatia (HR), Italy (IT), Lithuania (LT), Netherlands (NL), Poland (PL), Sweden (SE), Slovenia (SI), United Kingdom (UK).

**Which languages?**
English, French, Lithuanian, Croatian, Dutch, Polish, German, Danish, Slovenian, Czech, Spanish, Swedish, Finnish, Italian, Bulgarian, Romanian, Maltese, Hungarian, Portuguese, Estonian, Greek, Latvian, Slovak, Irish.

**What time span?**
Documents dated 1801-06-10 to 2026-11-12.

**Is anything missing?**
National case law (3,496,146 decisions) is not in these files: it is delivered separately, only after personal-data masking. United Kingdom: case law (Find Case Law) is excluded: UK case law (Find Case Law) is not offered: its licence does not permit computational analysis. Version history starts when a document first entered the corpus.

**Does it contain personal data?**
Legislation contains little personal data. CJEU case law is published by the Court with parties anonymised where the Court decided so. National case law is not included in this snapshot. Raw upstream HTML is never included.

**Is there a label or target?**
No labels. Citation edges and versions are structural facts from the sources, not annotations.

## Collection process

**How was the data acquired?**
From official upstream publications: EUR-Lex and the Cellar repository of the Publications Office of the European Union, and the official journals, legal databases and court portals of each national lane, through their APIs and open-data dumps.

**Over what timeframe?**
Continuously since the LexAPI corpus started; each version records the time it was observed (fetchedAt).

**How are source terms handled?**
Each lane’s upstream terms and access channel are documented per source and are under legal review. Documents a source lists for removal, and documents it withdraws, are removed and delivered as tombstones.

## Preprocessing

**What was done to the raw data?**
HTML, XML and PDF renditions are parsed into titles, metadata, recitals and articles; whitespace is normalised; a SHA-256 content hash is computed over the parsed text. Citations are resolved into typed edges only when the match is deterministic and unique.

**Is the raw data saved?**
Raw upstream pages are retained internally but never shipped.

## Uses

**Intended uses**
Retrieval-augmented generation, fine-tuning, pre-training (Lab tier), evaluation and benchmarking of legal AI systems; analytics on legislation and its citation structure.

**Uses to avoid**
Treating consolidated texts as authentic law; profiling judges or court clerks (prohibited for French data by art. L111-13 COJ); re-identifying anonymised or pseudonymised persons.

## Distribution

**How is it distributed?**
Parquet (zstd, ~512 MB shards) and gzip JSON Lines, with MANIFEST.json (counts, SHA-256 per shard), DATASHEET.md and LICENCES.md, by signed download or to the licensee’s bucket.

**Under what licence?**
The Arieta Data master licence with tier schedules (licence agreement on request). Upstream attribution passes through on every record. The public sample keeps each record’s upstream licence; Arieta’s structure and metadata in the sample are CC BY 4.0.

## Maintenance

**Who maintains it?**
Arieta Labs Ltd. Contact info@arietalabs.com.

**How is it updated?**
Monthly delta snapshots with upserts and tombstones. Licensees must delete tombstoned records.

**Where are changes announced?**
On the changelog (https://data.arietalabs.com/changelog/) and its RSS feed.