Provenance & AI Act
Where every record comes from, and what we leave out.
Arieta Data is compiled from official public-sector publications through a read-only export pipeline. Every record names its source and licence; withdrawn documents are tombstoned; raw upstream HTML is never shipped.
Last updated
-
01Upstream
Official publications
EUR-Lex, Cellar and the Official Journal for EU law; national official journals and legal databases for each lane. Publishers’ documented APIs and open-data dumps first. Where a source’s robots.txt disallows crawling but the publisher offers a documented API or bulk channel, we use that channel and record the reason.
-
02Ingest
LexAPI corpus
Documents are fetched by the LexAPI lanes, parsed to articles and metadata, and hashed. A new content hash creates a new version. Withdrawals are recorded in one table, the source of truth.
-
03Snapshot copy
Reproducible numbers
Snapshots are cut from a point-in-time copy of the corpus, so every count in a snapshot’s manifest is reproducible.
-
04Export
Filter and shape
Withdrawn rows are dropped, raw HTML is never selected, held lanes are left out, and each record gets its source, licence and attribution. Output is streamed to Parquet and JSONL.
-
05Snapshot
Manifest and datasheet
Counts and a SHA-256 per shard go into MANIFEST.json; DATASHEET.md and LICENCES.md are generated from the same data. The snapshot is immutable.
-
06Deltas
Monthly updates
Each month: upserts for new and changed records, tombstones for withdrawals. Licensees apply both.
Licences
Upstream licences, lane by lane
Each lane’s licence is recorded from the publisher’s terms and documented per source. Lane pages carry the full wording and attribution text; the attribution file ships with every snapshot.
Personal data
How personal data is handled
National case law is delivered masked
National court decisions name or pseudonymise parties, and in several countries judges. They (3.4M+ decisions as of the November 2026 snapshot) are delivered only after personal-data masking, including masking of magistrate and clerk names for France.
Withdrawals become tombstones
When a source withdraws a document, for example after a personal-data takedown, LexAPI records the withdrawal once. The export drops it, and the next monthly delta carries a tombstone. Licensees must delete tombstoned records.
No raw HTML
Upstream pages contain material outside the parsed fields, including personal data and page chrome. Raw htmlContent and htmlMetadata are never exported in any product.
GDPR responsibilities
Each licensee decides its own purposes and is responsible for its own processing; the licence passes upstream restrictions on. Arieta Labs does not claim any certification. Open questions, such as how a model trained on court decisions is treated, are part of our legal review and are discussed openly with buyers.
AI Act · Article 53(1)(d)
Training-content summary: paste-ready text
Providers of general-purpose AI models must publish a sufficiently detailed summary of the content used for training, using the AI Office template (Regulation (EU) 2024/1689, Art. 53(1)(d); applicable from 2 August 2025). The text below describes this dataset with the figures of the current snapshot. Adapt it to the template’s sections; it is not legal advice.
Which section? In the AI Office template, this dataset belongs in Section 2.2, private datasets obtained from third parties. Whether it is 2.2.1 (licensed from rightsholders) or 2.2.2 (other third parties) is a judgement call for your counsel; the guide sets out both readings. Name the licensor, the data type, the size and the time span.
Numbers update with each snapshot. The figures come from snapshot 2026-11. Regenerate the text when you train on a newer snapshot.
Read more: AI Act Article 53 training-data summary: a template for legal corpora.
Primary source: Regulation (EU) 2024/1689, Article 53.
Licensed dataset: Arieta Data, from Arieta Labs Ltd (England & Wales company no. 16629785), snapshot 2026-11.
Data type: text (legislation and court decisions), structured at article level, with metadata, version history and citation edges.
Size: 419,355 documents, approximately 4,070,000 articles, 15,098 document versions, 801,149 citation edges, approximately 2,050,000,000 tokens (characters/4, estimated from a sample), as of the November 2026 snapshot.
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.
Time span of content: 1801-06-10 to 2026-11-12.
Source category: private dataset obtained from a third party (Arieta Labs Ltd), compiled from publicly available official publications:
- European Union (EU): EUR-Lex legal notice (reuse authorised, source acknowledged; Commission documents under Decision 2011/833/EU)
- Austria (AT): CC BY 4.0 (RIS OGD, Namensnennung 4.0 International)
- Czech Republic (CZ): NKOD podmínky užití: no copyright work, no database right, no sui generis right (≈ CC0)
- Germany (DE): Official works outside copyright (§ 5(1) UrhG); no licence stated by GII / RII
- Denmark (DK): Betingelser for anvendelse af Civilstyrelsens åbne data (gratis, ikke-eksklusiv og ubegrænset brugsret; ophavsretslovens § 9)
- Spain (ES): Condiciones de reutilización de la AEBOE (licencia tipo, Resolución de 27 de junio de 2024; Ley 37/2007)
- Finland (FI): Creative Commons Attribution 4.0 International (CC BY 4.0) — Finlex open data
- France (FR): Licence Ouverte 2.0 (+ Judilibre CGU for case law)
- Croatia (HR): Official texts outside copyright (ZAPSP NN 111/2021 čl. 18 st. 3); odluke.sudovi.hr Uvjeti korištenja (automated download, any use, attribution)
- Italy (IT): CC BY 4.0 (Normattiva open data, in force from 1 Jan 2026; all records fetched in 2026; OpenGA: CC BY 4.0)
- Lithuania (LT): Creative Commons Attribution 4.0 (data.gov.lt datasets 2613 and 1938); legal acts and court decisions outside copyright (Autorių teisių įstatymas, art. 5)
- Netherlands (NL): No copyright on statutes and judgments (Auteurswet art. 11) and no database right (Databankenwet art. 8); Rechtspraak open data Public Domain Mark / CC0
- Poland (PL): No licence stated; official acts and judgments excluded from copyright (art. 4 Prawo autorskie), re-use under the Open Data Act 2021
- Sweden (SE): Riksdagens öppna data: free use and redistribution with attribution ("Källa: Sveriges riksdag"); court decisions and statutes exempt from copyright (9 § URL 1960:729)
- Slovenia (SI): CC BY 4.0 (Priznanje avtorstva), OPSI records for PISRS NPB and Sodna praksa
- United Kingdom (UK): Open Government Licence v3.0
Collection: Arieta Labs collected the upstream data through publishers' documented APIs, open-data dumps and official websites. Documents that a source lists for removal (for example takedowns listed in robots.txt) and documents withdrawn by the source are removed and delivered as tombstones. A per-source record of terms and access channels is supplied with each snapshot.
Personal data: national court decisions are not in this snapshot’s files; they are delivered separately, only after personal-data masking. CJEU case law is included as published by the Court. Documents withdrawn by their source are delivered as tombstones and deleted by the licensee.
Copyright and reuse: each record carries its source, licence and attribution. Reuse of EUR-Lex content is permitted under the EUR-Lex legal notice (Commission reuse policy, Decision 2011/833/EU); CJEU case law under the CURIA copyright notice; national texts under the terms listed per source. UK case law is not included: the Open Justice Licence does not cover computational analysis.
More: https://data.arietalabs.com/provenance/Need more for your legal review?
Ask for the licence agreement, the attribution file (LICENCES.md) for your lanes and the manifest of the current snapshot. We send them with a sample cut.
- We reply within one business day
- Custom cuts in 5–6 business days
- Free sample, no sign-up