Legislation as a Temporal Formal System: Evidence from 53,223 Versions of the Austrian Federal Statute Book

This dataset is a point-in-time, versioned representation of Austrian federal legislation, reconstructed from the consolidated texts published in the Austrian Legal Information System (RIS). It accompanies the paper Legislation as a Temporal Formal System: Evidence from 53,223 Versions of the Austrian Federal Statute Book. The corpus covers 15,209 instruments (federal statutes and ordinances) across 53,223 dated versions. Each instrument is stored as an independent git repository whose commit history is analogous to its amendment history: every consolidated version is one commit, dated by its entry into force, so any section can be retrieved as it stood on any date. Contents austrian-federal-law-git-corpus-complete.tar.gz — One git repository per instrument (15,209 repositories); the authoritative version history versions_fulltext_clean.jsonl.gz — Cleaned full text, one JSON record per version versions_fulltext.jsonl.gz — Raw full text, one JSON record per version austrian-federal-law-deposit.tar.gz — Documentation and tabular exports, a datasheet (Gebru et al. structure), versions.csv and instruments.csv, supporting RIS metadata, the coverage audit, the reconstruction and verification scripts, and a SHA-256 manifest (checksums.txt) The paper's published figures reproduce from this deposit. code/verify_all_claims.py checks 42 figures against versions.csv and instruments.csv (corpus counts, the full Table 1, amendment depth, the date conventions, the electoral cycle, the extraction and repair rates, and the two worked amendment examples) and exits non-zero on any mismatch; code/reproduce_electoral.py recomputes the electoral figures with binomial p-values; code/reproduce_findings.py recomputes the survival and temporal-machinery figures from the full-text export. Two figures, instrument coverage (98.3%) and version recall (98.9%), were measured by querying RIS at harvest time and are documented in audit/ rather than being recomputable offline. The commit_sha column in versions.csv cross-references each version to its originating git commit. Derived from RIS consolidated federal law, published as Austrian Open Government Data under CC BY 4.0. The authentic source of Austrian federal law is the Federal Law Gazette (Bundesgesetzblatt, BGBl), referenced per version in the data. This deposit is released under CC BY 4.0. For what purpose was the dataset created? Consolidated legal databases present each statute as a single "current" text. That presentation hides the temporal structure of the law: every section has a validity interval, amendments open and close those intervals, and a consolidated reading is only ever a snapshot. This corpus was built to make that temporal structure explicit and machine-readable, so that the evolution of federal legislation can be studied as a formal system rather than as a sequence of disconnected snapshots. It supports point-in-time legal question answering, retrieval-augmented generation over the correct historical text, and quantitative study of legislative change (amendment frequency, section lifetimes, structural volatility). Who created the dataset? The corpus was assembled by Dionis Ramadani as part of doctoral research on the temporal structure of federal legislation. The underlying legal texts are the intellectual and administrative product of the Republic of Austria, published through RIS by the Federal Chancellery (Bundeskanzleramt). Who funded the creation of the dataset? The reconstruction was carried out in the course of academic research. The source data is public administrative data provided by the Republic of Austria at no cost. Composition What do the instances represent? Each instance is one dated version of a federal instrument: the full text of a statute, constitutional act, or ordinance as it stood on a specific effective date (fassung_vom). Versions are grouped into instruments (one per legal act) and are stored so that the change between any two consecutive versions is directly recoverable. How many instances are there in total? The corpus contains 15,209 instruments comprising 53,223 dated versions (plus 4 versions carrying no parseable effective date, for 53,227 total stored commits). Of the instruments, 2,910 are statutes and federal constitutional acts (typ BG) and 12,299 are ordinances (typ V). The split is taken from the RIS consolidated-law type census: an instrument is classified V if its Gesetzesnummer appears in the RIS Typ=V enumeration, and BG otherwise. Federal constitutional acts are grouped under BG because RIS does not label them separately in a reliable way. Constitutional acts are covered only in part, and the corpus should not be read as covering Austrian constitutional law. RIS lists 99 federal constitutional acts; this corpus holds 29. The 70 absent (listed in audit/ris_missing_instruments.csv) include the Staatsgrundgesetz, the Habsburgergesetz, the Nationalsozialistengesetz and the 1925 recension of the B-VG, and 34 of them are still in force. The B-VG itself is present with its full version history; what is missing are the separate constitutional acts alongside it. Instrument coverage was measured by enumerating the RIS consolidated-law collection in full (358,397 documents scanned). The register holds 15,467 distinct federal instruments, of which this corpus holds 15,208, or 98.3 percent. Of the 259 absent instruments, 171 are already repealed and the remaining 88 are still in force; cutting across that split, 115 carry state-law titles misfiled in the federal collection. What data does each instance consist of? Each version records: the instrument's RIS abbreviation and Gesetzesnummer (RIS instrument identifier), the instrument type, the effective date, the promulgating gazette reference (Bundesgesetzblatt / BGBl citation), and the full sectioned text. Each section carries a stable section identifier, a heading, a structural type (for example Paragraf, Artikel), and the body text. Is any information missing? RIS does not assign an official short abbreviation to every instrument. For 2,930 instruments an official RIS abbreviation is used as the directory key; for the remaining 12,279 the key is a slug derived deterministically from the instrument's official title. This is a naming convenience only and does not affect the legal text. Four versions carry no machine-parseable effective date in their gazette reference and are retained without a date. Are relationships between instances made explicit? Yes. Within an instrument, versions are ordered by effective date and stored as successive commits in a git repository, so the diff between consecutive versions is exact and recoverable. Across instruments, the shared Gesetzesnummer and gazette references link related acts and their amendments. Are there recommended data splits? No. The corpus is a complete census of a legal system's federal holdings, not a sampled dataset, and is not intended for train/test partitioning. Studies that require held-out material should split by instrument or by time period explicitly. Are there any errors, sources of noise, or redundancies? The corpus inherits any transcription characteristics of the RIS consolidated texts. Section segmentation follows the RIS document structure. A small number of instruments carry sentinel or pro-futuro effective dates: 153 versions were dated in the future at the snapshot date of 2026-07-18 (pro-futuro entries with a known future commencement, for example the 2028 recodification of the ABGB), and 35 versions carry the RIS sentinel date 9000-01-01 used for provisions whose commencement is not yet fixed. Being future-dated is relative to the reading date: two of the 153 commenced on 2026-08-01, so 151 remain future-dated as of 2026-08-18. These are retained deliberately and flagged by their dates rather than removed. One source of noise deserves specific attention because it affects text-difference measurements. The corpus was built by two harvest generations that serialise identical legal text slightly differently: the later generation keeps the leading section marker in the body, uses non-breaking spaces around paragraph numbers, orders annotation blocks differently, and retains enumeration markers the earlier generation strips. Comparing a version from one generation against a version from the other therefore reports a difference in most sections even where the legal text is unchanged. Every row in versions.csv carries a harvest_generation column (original or reharvest_2026) so analyses can be restricted to one generation. Studies of coverage, depth and validity dates can safely use all versions; studies that diff consecutive texts should not mix generations. Is the dataset self-contained? Yes. All text is included; no external resources need to be fetched to read any version. The gazette references point into the Bundesgesetzblatt for provenance but are not required to use the corpus. Does the dataset contain confidential or sensitive data? No. It contains only published, in-force or historical legal texts of the Republic of Austria, which are public by their nature. It contains no personal data beyond what appears in the published statutory text itself. Collection process How was the data acquired? Versions were reconstructed from the RIS consolidated federal law application ("Bundesrecht konsolidiert") through the official RIS OGD interface. For each instrument, the set of consolidated versions was enumerated and each version's structured document (NOR-XML) was retrieved from the RIS document endpoint, parsed into sections, and committed to a per-instrument git repository with the effective date and gazette reference recorded in the commit metadata. What mechanisms were used to collect the data? Retrieval used the RIS OGD API (data.bka.gv.at/ris/api) for enumeration and the RIS NOR-XML document endpoint for full text, over HTTPS, with a conservative request rate. No scraping of the HTML front end was involved. Over what timeframe was the data collected? The corpus was assembled in 2026 and reflects the state of the RIS consolidated holdings at that time. The legal texts themselves span effective dates from 1772 to 2035 (the latter being pro-futuro commencements already enacted). Were any ethical review processes conducted? Not applicable. The source material is public legislation. Preprocessing Was any preprocessing done? Yes. Duplicate instrument records arising from overlapping RIS enumerations were removed by deduplicating on the Gesetzesnummer, so each instrument appears once. Versions sharing an identical effective date that arise from the same consolidation event were collapsed to a single stored version. Section text is preserved as delivered by RIS; no normalization of legal wording was applied. In addition, a harvesting defect was identified and remediated after the initial export. In 371 instruments, 53,707 stored section bodies (1.5% of the 3,472,526 section instances, spread over 4,124 consolidations) had accumulated duplicated text: either the same consolidated text stored twice in succession, or the machine-readable text concatenated with the RIS read-aloud rendering (in which references are spelled out, e.g. "Paragraph 5, Absatz zwei"). These were repaired in versions_fulltext_clean.jsonl.gz under a fixed precedence. First, collapse of a provable repetition (39,105 instances): a block is deleted only when its canonical token stream equals the block immediately before it, under a canonicalization that maps the read-aloud rendering and machine citation form onto one token stream (Paragraph 5, Absatz eins, Ziffer eins and § 5 Abs. 1 Z 1 canonicalize identically; a letter-suffixed section number is split so that § 159a and the read-aloud Paragraph 159 a, also agree), and only when the second block either carries read-aloud vocabulary or is byte-identical to the first. The retained characters are always a subsequence of the stored body, so this step cannot introduce text that was not already present; the property is asserted on every repair. Collapse is iterated to a fixpoint, because a body carrying three copies of the same block needs two removals. Second, where collapse did not resolve the body, re-fetch of the authoritative section text from RIS (14,445 instances), accepted only under an anachronism guard: since the defect only adds words, a replacement is rejected if more than 5% of its canonical vocabulary is absent from the stored body, which prevents a later consolidated text from silently replacing a historical version. 157 sections resisted both methods and are shipped unaltered; together with three re-fetches that did not change the body this leaves a flagged residual of 160 section instances (0.005%). All 160 are listed with their full text in residual_flagged.csv. Every repair is recorded in repair_log.csv with method and before/after body lengths. The pre-remediation export is retained verbatim as versions_fulltext.jsonl.gz for auditability. The defect affects only the full-text export; versions.csv and instruments.csv are unaffected. After remediation, 99.8 percent of section instances carry a non-empty body. Was the raw data saved? The reconstruction is reproducible from RIS. The per-instrument git repositories are the primary artifact and preserve every retrieved version verbatim. Is the preprocessing software available? The reconstruction pipeline (git-for-law-austria) that retrieves and parses NOR-XML is part of the associated research project; the release and repair scripts are included in code/. Uses What has the dataset been used for? Quantitative analysis of the temporal structure of Austrian federal legislation: section-level validity intervals, amendment frequency and section lifetimes, structural volatility of the legislative "machinery," the prevalence of pro-futuro and open-ended commencement dates, and the distribution of commencement dates across the parliamentary term. What other tasks could it be used for? Point-in-time legal question answering; retrieval-augmented generation constrained to the text in force on a given date; training and evaluation of temporal legal reasoning; comparative legal-informatics studies of legislative change. Is there anything about the composition or collection that might affect future uses? Three points matter. First, the consolidated texts are non-authentic: under Austrian law the authentic promulgation is the Bundesgesetzblatt, and RIS consolidations are informational and carry no legal force. Any use that requires the legally binding text must consult the BGBl citation recorded with each version. Second, the corpus covers federal instruments with a RIS consolidated entry; it does not include Land (state) law, EU law as applied in Austria, or case law, and it covers constitutional acts only in part (see Composition). Third, version recall is lower for the most-versioned instruments (see Composition), so their per-version counts are lower bounds. Are there tasks for which the dataset should not be used? It must not be used as an authoritative source of the binding legal text (see non-authenticity above), and it should not be presented as legal advice. Distribution How is the dataset distributed? As a Zenodo record (DOI 10.5281/zenodo.21855630) of four files: austrian-federal-law-deposit.tar.gz (this documentation, the tabular exports versions.csv and instruments.csv, the supporting RIS metadata, the coverage audit under audit/, the reconstruction and verification scripts under code/, the datasheet, the licence, and a SHA-256 manifest covering the payload); versions_fulltext_clean.jsonl.gz (the remediated full text, recommended); versions_fulltext.jsonl.gz (the pre-remediation full text, retained for auditing the repair); and austrian-federal-law-git-corpus-complete.tar.gz (the 15,209 per-instrument git repositories, which carry pre-repair text). The commit_sha column in versions.csv is the SHA of the originating commit in that corpus, so the flat export and the git history are mutually addressable. The published figures reproduce from the flat files via code/verify_all_claims.py, code/reproduce_electoral.py and code/reproduce_findings.py, each of which asserts every number it prints and exits non-zero on a mismatch. Under what license? The RIS consolidated federal law is published as Austrian Open Government Data under Creative Commons Attribution (CC BY 4.0). This derived corpus is distributed under the same license, and the code under code/ is released under the MIT license. Attribution is due to the Republic of Austria / RIS as the source of the legal texts. See LICENSE. Are there IP or other restrictions? Beyond the CC BY attribution requirement, none. Users must not represent the consolidated texts as legally authentic. Maintenance Who maintains the dataset? Dionis Ramadani. How can the dataset be updated? The corpus is a snapshot. Because the reconstruction pipeline is deterministic against RIS, a later snapshot can be produced by re-running enumeration and retrieval; new versions append as new commits without disturbing history. Will older versions continue to be supported? The released snapshot is immutable and identified by its SHA-256 manifest. Future snapshots are additive.