Which Legal Artifacts Belong in the Twin

Anyone who turns a statute into a digital twin faces a question on day one that no architecture diagram answers: which legal sources go in at all? Only the text of the statute, or also the ordinances that fill it in, the court rulings that interpret it, the commentaries that explain it, the constitution and EU law that limit it? This fourth post in our series on digital twins of legislation proposes a simple ordering principle: what decides a source's place in the twin is not its rank but its sphere of effect, that is, whether it is binding, interpretive, or informative.

A statute never comes alone

In the first three posts we described what a statute can do as a model, how prose becomes an executable twin in four layers of text, ontology, configuration, and logic, and why the municipal tax is a suitable test case. In all three posts, "the statute" stood at the center as if it were a single document. In practice, it never is.

To calculate the Upper Austrian tourism levy, you need the Upper Austrian Tourism Act 2018, plus the ordinances of the provincial government that place municipalities in municipality classes and assign businesses to contribution groups, and you should know how the administrative courts have understood individual terms. Roughly 29,000 levy payers have to reproduce this calculation in their own systems (Schnitzhofer 2025, ESWC PhD Symposium). For the municipal tax, it is the Municipal Tax Act 1993, the fiscal constitution and the Fiscal Equalization Act that allocate the tax to municipalities, decrees of the Ministry of Finance, an extensive body of Supreme Administrative Court case law on terms such as permanent establishment and employee, and several commentaries. Above all this sit EU law and fundamental rights, which determine what an authority may decide automatically in the first place.

Since Kelsen and Merkl, legal scholarship has ordered these sources in the hierarchy of norms: constitution above statute, statute above ordinance, ordinance above individual administrative act. The hierarchy tells us which norm prevails in a conflict. It does not tell us what occupies us when building the twin: what does an artifact do to the model once we let it in?

Three spheres of effect: binding, interpretive, informative

In our research at the Institute of Business Informatics – Data & Knowledge Engineering at JKU Linz, we use a working classification for this, which we call the sphere of effect. It asks not about rank but about the effect an artifact has on the outcome of a decision:

• Binding. The artifact creates the norm itself. Without it, the obligation, the tax rate, or the exemption would not exist. This includes the constitution, directly applicable EU law such as EU regulations, federal and provincial statutes, and the ordinances issued on their basis. A ruling of the Constitutional Court that strikes down a provision also belongs here, because it changes the binding text.
• Interpretive. The artifact creates no norm but determines how its terms are to be read. This includes the case law of the supreme courts and of the Court of Justice of the European Union, the explanatory materials of bills, decrees and guidelines of the ministries, and the commentary literature. Decrees are an instructive borderline case: they bind subordinate authorities, but neither citizens nor courts. For the twin, they are therefore interpretive, not binding.
• Informative. The artifact explains without binding or interpreting: information brochures, questions and answers on government websites, calculators, textbooks, journal articles, and not least the general knowledge of a language model. Such sources help to understand the use case and to sanity-check results. They may never justify a rule.

The value of this classification: an artifact's sphere determines which layer it belongs to and what role it plays there.

What binds the twin is not the same as what the twin executes.

Binding sources: the anchor in the text layer

For binding sources, one rule applies without exception: every rule in the logic layer and every value in the configuration layer refers to a passage in a binding source, with an identifier and a version date. The legal text remains the primary legal source; the twin is its computing companion, pointing back at every decision to the provision from which it follows. The European Legislation Identifier (ELI) serves this purpose well; it makes legal provisions in Austria and in the European Union addressable down to the level of individual sections, and the Austrian Legal Information System (RIS) provides it for federal and provincial law.

For the municipal tax, the binding parameters are manageable: the tax rate of 3 percent of the assessment base, the allowance of 1,095 euros, and the threshold of 1,460 euros per calendar month are laid down in § 9 of the Municipal Tax Act 1993; the catalog of exemptions is in § 8, and the rules for apportioning the assessment base among several municipalities are in § 10. All of these values and lists belong in the configuration layer, each with its section as an anchor. If an amendment changes the allowance, a value in the configuration changes, not the code. For the tourism levy, the percentage rates per contribution group and municipality class are laid down in the statute itself (§ 43 of the Upper Austrian Tourism Act 2018); which rate applies to a given business, however, is decided only by the ordinances on municipality classes and contribution groups. Formally they are not statutes, but they are binding, and therefore anchors as well.

An ordinance ranks below the statute in the hierarchy of norms, yet it often carries the classifications and numbers without which the twin cannot compute. Whoever models only "the statute" ends up with a twin whose parameters are incomplete.

Cross-references to other statutes are a special case. The Municipal Tax Act defines the employee by reference to the Income Tax Act; the tourism levy ties in with revenue concepts regulated elsewhere. Such provisions are binding but belong in the twin of the other statute; we reference them instead of copying them. In this way, individual twins gradually grow into a system of systems, as digital twin research describes it (Schnitzhofer 2025, ESWC PhD Symposium).

Constitution and EU law: frame, not computational core

The constitution and EU law are binding, but they rarely contain anything the twin could compute. The fiscal constitution and the Fiscal Equalization Act say that municipalities levy the municipal tax; they do not say how high it is. The General Data Protection Regulation regulates in Article 22 the conditions under which a decision may be made solely by automated means. The AI Act, Regulation (EU) 2024/1689, in force since August 1, 2024, exempts in Recital 12 software that automatically executes rules defined solely by humans, and lays down logging and oversight obligations for high-risk systems.

Such norms belong in the twin, but in a different place than the tax formula. They are norms the twin is subject to, not norms the twin executes. Their place is the metadata and the interface definitions: which authority is competent, on which basis of competence, which decision fragments may be answered without human review, which logs must be produced. Where EU law exceptionally contains computable rules itself, for instance thresholds in a directly applicable EU regulation, it changes role and becomes an anchor in the configuration. The sphere stays the same; the layer changes.

Rulings and commentaries: interpretation moves into the ontology

Interpretive sources are the most demanding, because they change the twin without a single letter of the statute changing. The Municipal Tax Act counts managing directors with a substantial holding in a corporation among the employees; when they are actually integrated into the operational organization has been specified by the Supreme Administrative Court in a long series of rulings. The statute names the group of persons, the criteria come from case law, and they decide the assessment base. When a commentary explains which activities fall under the concept of entrepreneur in the Tourism Act, that hierarchy belongs in the ontology.

The ontology layer defines the terms, their subclasses, and their relationships; every definition that is not taken verbatim from the statute carries an annotation with source, date, and a note that it is interpretation. The second home of interpretation is the test cases: a ruling with a clear set of facts is an excellent test case with an expert-validated expected outcome. Case law thus becomes verifiable instead of vanishing invisibly into the code.

Two limits must be observed. First, an interpretive source may never be the only anchor of a rule; the rule refers to the statute, the interpretation sharpens the term. Second, selective automation applies, which we have formulated as the first of four feasibility conditions for automation that complies with the rule of law: where a term is open, case law inconsistent, or discretion provided for, the twin exposes no interface but hands the case to a human. A twin that falls silent on contested interpretation is not a defect but a feature.

An organizational duty follows from this: whoever operates a twin monitors case law just as closely as the Federal Law Gazette. If a supreme court changes the understanding of a term, the ontology needs a new version, and the test cases must show it. The twin is living legislation that grows with interpretation.

Informative sources stay outside

For informative sources, the rule is the shortest: they stay outside the twin. Brochures, lists of questions, textbooks, and the knowledge of a language model help to understand the use case, delimit the scope, and check results for plausibility. But as soon as a rule or a parameter in the twin is supported only by such a source, the model has a defect.

The handling of language models shows this most clearly: they are well suited to proposing candidates for terms, parameters, and rules from the text, provided humans confirm or correct every artifact; as a legal source, they are useless. A model that estimates the tourism levy from its general knowledge delivers a plausible number and no justification; a model that asks the twin delivers the amount and the section. Informative sources may enrich the explanation, never carry the decision.

Five rules for selection

From the three spheres follow five selection rules, which we apply in our use cases:

1. Every rule has a binding anchor. Logic and configuration refer exclusively to a statute, an ordinance, or directly applicable EU law, with identifier and version date.
2. Interpretation is annotated, not hidden. Rulings, explanatory materials, decrees, and commentaries shape the ontology and the test cases and are marked there as interpretation, with source and date.
3. The frame lives in the metadata. Constitution, fundamental rights, GDPR, and AI Act determine competence, limits, and clearance for automation, not the formula.
4. Informative material stays outside. Brochures, FAQ, textbooks, and model knowledge serve understanding and plausibility checks, never justification.
5. Scope follows determinacy. Where discretion, open terms, or inconsistent case law prevail, there is no interface; the case goes to a human.

Whoever applies these rules gets a twin whose provenance is legible in every line. That is exactly what the rule of law demands of an automated decision: not only the correct result, but a statement of which norm it follows from and which interpretation underlay it.

The next step

How the four layers and the spheres of effect interact in our research program is described on the Digital Twin of Legislation page. The work published so far, including the contributions to IRIS 2025, the ACM Symposium on Computer Science and Law 2025, and the ESWC 2025 PhD Symposium, can be found under Publications. If you want to model a statute in your administration or your organization, start with the list of artifacts and their spheres; we are happy to discuss your use case.