Digital Twins of Legislation: Where the Research Stands Today

Administrative law is published as prose. Every authority, every software vendor, and every business translates the same provisions once again into its own program code, expensively, repetitively, and without the decisions made along the way being documented. Since 2024, we have been working at Johannes Kepler University Linz, together with ReqPOOL, on mechanizing this translation without giving up the guarantees of the rule of law. This article describes what the research is working on as of August 2026: an agentic pipeline that turns statutory texts into verifiable digital twins of legislation, and a benchmark that measures every design decision against the available alternatives. The results of this work will appear with the scientific publication; this article is about the research question, the structure, and the current state of the project.

The problem: law is translated into code thousands of times over

In 2020, Dag Wiese Schartum described how legal sources become program code. The first step, the qualification of the legal sources, is almost always carried out implicitly and recorded nowhere. Rules as Code approaches such as OpenFisca, Catala, or PROLEG have made the product executable; the translation itself remained manual work. The consequence is divergent implementations: identical facts produce different results in different systems. That is first a problem of the rule of law and only then a software problem.

Three premises from the program's previously published studies outline the task:

• Economic. Decentralized translation is a first-order cost factor. For the Upper Austrian tourism levy, with around 29,000 affected businesses, the burden study (JURISIN 2026) shows that a centrally maintained, machine-executable twin reduces the administrative burden of implementing the law roughly 35-fold compared with decentralized implementation. For comparison: the PROLEG formalization of the contract law part of the Japanese Civil Code took around 900 person-days of expert effort over ten years; the German National Regulatory Control Council puts the compliance costs of new federal regulation for 2025 alone at 17.3 billion euros.
• Normative. The requirements for automation that conforms to the rule of law are known and strict. A Grounded Theory study with nine senior experts from EU legislative bodies, a national parliament, legislative drafting, the notarial profession, public-sector software, and academia (JURIX 2025, AIOG 2026) yielded four feasibility conditions: only deterministic partial decisions with little discretion are automated; legal semantics are standardized and ontologically aligned; norms are transferred into computable structures that keep every output traceable to the promulgated sources; and legislative and governance processes must keep the text and its executable counterpart synchronized.
• Technological. Unsecured language models satisfy the economic premise and fail the normative one. The literature reports 58 to 88 percent hallucinations for bare models on verifiable legal questions and 17 to 43 percent for commercial retrieval-backed legal tools. Where the law demands a specific result, probabilistic answering is inadmissible, and every explanation must cite the governing provisions.

For every approach deployable today, these three premises are jointly unsatisfiable. This is precisely where the research question comes from: Which architectural commitments and which construction process make machine-generated executable law norm-conformant by design, and which of these commitments are necessary because without them a known failure mode returns?

What a digital twin of legislation is

A Digital Twin of Administrative Law (DTAL) consists of four deliberately separated layers, which we have described in the program's previous contributions:

1. Statutory text. The promulgated wording of every selected legal source, verbatim and unaltered, each passage with citation and document class.
2. Ontology. A machine-interpretable vocabulary for subjects, concepts, and parameter classes, in OWL and RDF.
3. Configuration. A machine-readable graph with the interface and every normative value, such as rates, thresholds, and allowances, each with a reference to the governing provision.
4. Logic. An executable decision function that maps case data, ontology, and configuration to a decision and explains it by reference to the norms.

The separation is the core of the approach: the text remains authentic, the ontology fixes the vocabulary, the configuration carries everything an amendment can change, and the logic carries only structure. From this follows the goal against which the design must be measured: if the legislature changes a rate or a threshold, the behavior of the published twin should change through an entry in the configuration, not through new program code. And because a call consists of machine-readable case facts, no decision should depend on how a case was described in natural language.

Published twins are versioned and provided through two interfaces: a web interface (REST) for e-government applications and enterprise software, and a tool interface for AI agents via the Model Context Protocol. A twin is thus a tool for conventional software and for agentic AI at the same time, and no language model stands in the decision path.

The anchor case: the municipal tax

The current anchor case is the Austrian municipal tax (Kommunalsteuer) under the KommStG 1993, regulated uniformly at the federal level, administered by the municipalities, and computable to the cent. Businesses pay 3 percent of the monthly gross payroll per municipal permanent establishment (§ 9). If the monthly assessment base does not exceed 1,460 euros, an allowance of 1,095 euros is deducted. Added to this are exemptions (§ 8), the six-month limit for temporary agency work (§ 7), and the apportionment across several municipalities (§ 10). Legally, the case is demanding: competence provisions of the constitution, referenced income tax and social insurance statutes, EU directives on temporary agency work, case law, and ministerial decrees belong to it, while lexically similar distractors such as the Vienna employer levy (Dienstgeberabgabe) must stay out. Precisely this mixture of determinism and legal depth makes the municipal tax a suitable test case. Research partners are the Federal Chancellery, the City of Linz (Office of the Magistrate Director), Johannes Kepler University, and Binary Growth GmbH.

What we are currently researching: the agentic pipeline

The twins built so far emerged from a procedure in which a language model proposes artifacts and experts validate every step. The ongoing work goes one step further and examines whether a staged pipeline of specialized agents can mechanize the translation to the point where central construction with N-fold reuse preserves the measured order-of-magnitude advantage while respecting every feasibility condition. Three research questions guide the work:

• Architecture. Which representation and which machine-checkable layer contracts make an executable replica of a legal regime traceable, equal in treatment, and maintainable under amendments by design?
• Generation. Which class of construction methods, from lexical retrieval through citation networks and graph-based selection to LLM agents, produces such representations reliably, under which conditions, and what do grounding in the sources, contracts, and human review each contribute?
• Dominance. Does agentic construction outperform manually created Rules as Code, unsecured model use, and retrieval-based procedures in computational correctness, citation completeness, and independence from case wording, while preserving the economic advantage?

The design under examination rests on three commitments. Grounding: agents propose legal sources exclusively via stable, resolvable identifiers, such as ELI addresses in the Federal Legal Information System (RIS), ECLI for case law, or Findok for decrees; full texts are retrieved, never generated, and an invented identifier does not resolve and therefore never enters the library. Contracts before trust: the layers are connected by machine-checkable rules, for example no normative value may appear as a program literal in the logic, and every decision must deliver a derivation log that names, for every parameter and every rule, the provision from which it originates. Human approval with version authority: agents produce proposals, review gates filter mechanically, and only acceptance by an author changes the project state; every version is immutable and attributable.

The pipeline works in stages that refine Schartum's transformation process: source proposal and corpus selection correspond to the qualification of the legal sources; passage selection with an assessment of twin-worthiness along the dimensions of addressability, normativity, computability, and rule-of-law determinacy, together with the generation of the ontology and configuration layers, correspond to interpretation and specification; the logic layer corresponds to formalization; verification against the confirmed specification and the approval workflow correspond to testing. The difference from practice: every step that Schartum locates in the minds of the experts is externalized as a persisted, versioned, and citable artifact.

How we evaluate

Whether the commitments are necessary can only be shown by measuring against the deployable alternatives. To this end, a source-traceable benchmark on the municipal tax is being built that turns corpus selection, the previously undocumented opening step, into a measurable stage. Compared are deterministic selection procedures, agentic variants, and model-based construction with and without grounding in a curated catalog, each in the same four-layer architecture and against the same expert-validated test cases with euro-exact reference amounts.

Measurement follows the design requirements derived from the preceding studies: authenticity of the statutory text, traceability of every normative value, computational correctness, completeness of the cited legal basis, equal treatment regardless of wording and repetition, maintainability under amendments, human oversight, explainability in legal terms, auditability, isolation of generated code, and economic viability. The approach is positioned against Rules as Code, against unsecured language models, and against procedures that search for evidence case by case. The figures, tables, and statistical tests of this evaluation will be published with the paper; until then they deliberately remain unpublished.

The limits of the project are stated: one regime, one jurisdiction, a European and expert-heavy sample in the qualitative study, a burden model based on expert estimates. Replication across further Austrian levies is prepared; the field comparison against a real conventional implementation project is the natural next study. Further axes are versioned resources and temporal test sets, formal verification of the generated logic against the configuration schema, and the use of twins by planning agents that compose several twins as tools.

Position within the research program

The work is the integration stage of a staged Design Science program. The first stage delivered the feasibility conditions and design principles with the Grounded Theory study, the second the four-layer architecture and the burden study on the tourism levy, the third the benchmark instrument for corpus selection. The fourth stage brings formalization, reference implementation, and evaluation together.

The preliminary work is published: the Grounded Theory study as a poster at JURIX 2025 (with Christoph G. Schuetz), the design principles at the AIOG Workshop of ICAIL 2026 and the burden study on the tourism levy at JURISIN 2026 (both Schnitzhofer, Nikiforova, Schuetz), the architecture at the SEMANTiCS 2025 Developers Workshop and at the NLL2FR Workshop of JURIX 2025. The book The Self-Driving State (Springer 2026) places the digital twin of legislation in the overall picture of the self-driving state. An overview of all contributions of the program can be found under Publications; the research roadmap to 2030 is in the article The Research Roadmap for Digital Twins of Legislation.

The next step

For public administrations, the state of the research means: the deterministic, parameterizable layer of administrative law, that is, levies, fees, rates, thresholds, deadlines, and eligibility checks, is the area in which digital twins of legislation can be centrally constructed, verified, and provided as a service, while open value judgments and discretion remain with humans. For businesses it means: a published twin can replace their own reprogramming of every amendment with a single call that returns a traceable derivation log.

If you want to examine which of your procedures or levies are suited to a digital twin of legislation, we are happy to discuss this in concrete terms: Book an expert meeting. How ReqPOOL supports public administrations along this path is shown on the Public Administration page.

Sources

• Florian Schnitzhofer, Christoph G. Schuetz, Enhancing Automated Decision-Making in Administrative Law Through Digital Twins of Legislation: A Grounded Theory Approach, JURIX 2025, IOS Press, pp. 436–438.
• Florian Schnitzhofer, Christoph G. Schuetz, Towards Translating Natural Language Normative Text into a Digital Twin of Administrative Law, NLL2FR Workshop, JURIX 2025.
• Florian Schnitzhofer, Christoph G. Schuetz, Towards a Scalable Architecture for Legal Ontologies Integrated into Digital Twins of Administrative Law, SEMANTiCS 2025 Developers Workshop.
• Florian Schnitzhofer, Nikiforova, Christoph G. Schuetz, Digital Twins of Legislation for Explainable Automated Decision-Making in Administrative Law, AIOG Workshop, ICAIL 2026.
• Florian Schnitzhofer, Nikiforova, Christoph G. Schuetz, Reducing Administrative Burden by Automating the Translation of Administrative Law into Digital Twins of Legislation, JURISIN 2026, LNCS, Springer.
• Dag Wiese Schartum, From Legal Sources to Programming Code, 2020.
• Florian Schnitzhofer, Patrick Pils, Philipp Seper-Ambros, The Self-Driving State, Springer 2026. About the book