From Research Framework to Platform

For eight posts, this series has been about how a statute becomes a digital twin: four layers, a translation path, four feasibility conditions, a burden model, and use cases ranging from the tourism levy to company formation. To close the series, I ask the question that ultimately counts in administrative practice: what separates a research framework from a platform that runs every day, and how much of that exists at the end of 2025? The answer has three parts: an open-source base, a first twin in public operation, and an operating model that governs versioning, certification, and responsibility.

Discarded variants and one framework: what has emerged by the end of 2025

The path to today's architecture led through discarded variants. A purely XML-based attempt in 2023 on a collective agreement missed both the automation and the cost targets. At IRIS 2025 in February, we derived from this the four-layer architecture of text, ontology, configuration, and logic and demonstrated it as a proof of concept on an excerpt of the Income Tax Act. A prototype in Prolog for the Upper Austrian landscape levy proved hard to grasp for legal practitioners and computer scientists alike; the tourism levy twin is therefore written in Python, with the ontology in OWL. In June, the PhD Symposium of ESWC 2025 brought the rule-of-law framing; in September, the Developers Workshop of SEMANTiCS 2025 brought the reference architecture that Christoph Schütz and I demonstrated on the twin of the Upper Austrian tourism levy (CEUR Vol. 4064). In December, JURIX 2025 added the four feasibility conditions from our grounded theory study with nine experts (Schnitzhofer and Schütz 2025, doi 10.3233/FAIA251630), and the NLL2FR workshop added the seven-step translation path from prose to twin.

The posts in this series have placed that in context: the four layers, the feasibility conditions, and the burden model, according to which a centrally maintained twin reduces the implementation effort for the tourism levy by roughly 35 times under the assumptions of our model. Together, this is a research framework: architecture, method, principles, and the evidence on a real statute whose twin returned the correct result in all 100 scenarios precalculated by a tax expert. It answers whether deterministic administrative law can be twinned; what remains open is how a twin is operated for everyone it affects.

What separates a prototype from a platform

A research prototype belongs to one team, models one functionality, and stays as it was published. A platform belongs to no one alone, holds many twins in every consolidated version, and is reachable through defined interfaces for the specialized applications of public administration, for business software, and for AI systems. The difference lies not in the twin but in operations.

In the reference architecture, the Digital Execution System takes on this role. It hosts any number of twins as container services, receives case data, invokes the logic of the matching twin, and returns the result together with its derivation log. Three kinds of clients connect: e-government applications and business software via REST, and AI systems via the Model Context Protocol, which exposes the REST interface as a tool for language models. For every twin, one rule applies: the configuration layer exposes an interface only for decision fragments that can be computed deterministically from the required inputs. Where discretion, open-textured legal terms, or missing evidence prevent that, there is no interface, and the case goes to a human.

The open-source base: why the code must be public

Openness has been part of the program from the start. At IRIS 2025 and at the ACM Symposium on Computer Science and Law in Munich in March, we named an open-source framework as a goal; at ESWC 2025, we announced design patterns and tool support under an open license. Since SEMANTiCS 2025, the source code of the tourism levy twin, including its REST and MCP interfaces, has been on GitHub, and the 100 test scenarios with the tax expert's expected values have been published as an open dataset on Zenodo. In our JURIX paper, we made a commitment: all artifacts of future use cases, that is, ontologies, executable rules, and step-by-step implementation guides, will be released under an open license; the first use case is publicly accessible at lawdigitaltwin.com; and every subsequent release will follow the same commitment.

This is not a gesture but a condition. A twin that calculates levies is the application of the law; whoever is not allowed to read it cannot verify that application. In my view, three reasons argue for it:

• Reviewability. Courts, audit institutions, affected parties, and their representatives must be able to trace which rule with which parameters led to a result. With open code, that is a matter of reading, not of trust.
• Equality before the law. If everyone uses the same implementation, there are no divergent interpretations between authorities, software vendors, and businesses. It is precisely these divergences that arise when every organization runs through the translation process from legal source to program code that Dag Wiese Schartum described in 2020 for the same statute once more on its own.
• No lock-in. The logic is defined by the statute, not by a vendor. Vendors can operate platforms, build integrations, and guarantee operations, but no one can own the rule.

For ReqPOOL, this is consistent. We do not sell licenses; our contribution consists of guiding statutes through the four layers, having the twins accepted by domain experts, and anchoring their operation in public administration and in companies.

lawdigitaltwin.com: the first twin in public operation

At lawdigitaltwin.com, the twin of the Upper Austrian tourism levy runs outside the laboratory. The platform lists the available twins in a searchable catalog, executes a twin in a form and shows the result with its breakdown, documents integration via REST and the Model Context Protocol, and offers the source code for download. The calculator derives the contribution from municipality, business activity, and the revenue of the year before last, and names the contribution group, municipality class, percentage rate, and minimum contribution from which it results. For the roughly 29,000 businesses in Upper Austria liable for the levy (estimate of the state government, Schnitzhofer 2025, ESWC PhD Symposium), this means that their accounting software, their tax advisor, or an assistant in a chat can use the same verified calculation.

Why the connection to AI systems matters is shown by our NLL2FR paper. A current language model, GPT-5 Pro with deep search, was asked to derive the levy directly from the statutory text: it gave incomplete or incorrect answers on several edge cases, overlooked an exemption for certain business types, and answered the same question differently across repeated runs. The twin returned the expert's result in all 100 scenarios. The Model Context Protocol combines both strengths: the model conducts the conversation, calls the twin as a tool, adopts its amount, and explains it with reference to the provisions. It no longer guesses and does not calculate itself; it translates.

A digital twin of legislation is a versioned, verified, and publicly inspectable artifact. Language models may query it; they cannot replace it.

Operating under rule-of-law conditions: three rules

A twin is only correct as long as it is synchronized with the statute, and only legitimate as long as someone stands behind it. We consider three rules indispensable.

Versioning: one statute, one version, one library

Every twin belongs to a Digital Twin Library that models exactly one statute in exactly one consolidated version. Name, jurisdiction, version, and date are part of the artifact, every provision is addressed through the European Legislation Identifier, and parameters carry their effective dates. An amendment creates a new version with a changelog, so that any decision can later be recalculated with the values that applied at the time of the case. The separation of layers pays off here: a changed percentage rate is a change in the configuration, while a new contribution group touches ontology and logic. The platform must distinguish both, log them, and synchronize them with an external version control system so that the history remains verifiable from the outside.

Certification before entry into force

A twin that calculates legal consequences must not go into operation silently. We propose treating it like an annex to the statute: it is drafted together with the text, checked before entry into force by domain experts, where appropriate with the involvement of the judiciary, against expert-validated cases, and then published as a verified version. For the transition period, one clear rule applies: where twin and wording diverge, the wording prevails until the twin is corrected. What legal status a certified twin receives, whether as an official annex, as a regulation, or through an enabling provision, is a decision for the legislature. We named formal certification pathways embedded in the promulgation process as an open research question at SEMANTiCS 2025; I consider it the most important institutional task of the coming years.

A designated body with a duty to keep it current

Computable structures help only if an organization has the duty and the capacity to update them with every amendment; that was the fourth feasibility condition of our interview study. Whoever publishes the twin is responsible for keeping it current, just as today a body is responsible for the consolidated version in the legal information system. This body needs lawyers, experts in the respective domain, and legal engineers, an official machine-readable vocabulary of legal terms to which new statutes refer, and metrics: administrative burden and error rates are monitored like financial costs. The AI Act, Regulation (EU) 2024/1689, requires logging, traceability, and human oversight for high-risk systems; a twin that logs every rule with its provision and explicitly excludes discretion meets these requirements by construction.

What comes next

The path from platform to regular operation is incremental. For a public administration, it begins with a single deterministic, parameterizable functionality with a high case volume: a levy calculation, a fee assessment, a deadline check. The twin initially runs in parallel with existing processing, its results are compared with the issued decisions, and only when they match does it take over routine cases. A company integrates the officially provided calculation through the interface instead of programming it and maintaining it with every amendment.

The research continues in three places. The next anchor case is the municipal tax under the Municipal Tax Act 1993, with its rate of 3 percent, the allowance of 1,095 euros, the exemption threshold of 1,460 euros, and the apportionment across several municipalities, together with the Federal Chancellery, the City of Linz, Johannes Kepler University, and Binary Growth GmbH. In parallel, the systematic study of legal ontologies from 2017 to 2025 is clarifying which existing ontologies can be integrated into the twin. And the translation path itself is to become tool-supported: language models propose segmentation, ontology classes, parameters, logic, and test cases, and experts confirm, correct, or reject each proposal before it enters the library. In our NLL2FR paper, we set down this division of labor as a principle: language models support drafting, while validation remains formal or expert-driven. The research prototype for this is being built and is not yet available to third parties. I consider the order decisive: the platform comes before the automation of translation, because only a platform with versioning and acceptance gives a model's proposals a place where they can be verified.

The next step

The vision behind this series is described on the page Digital Twin of Legislation; the papers for IRIS, SEMANTiCS, and JURIX 2025 can be found under Publications. Anyone who operates a provision in a public administration that is reprogrammed many times over today, or who maintains a calculation in a company that follows every amendment, can guide it with us through the four layers and put it into operation on a platform with versioning and acceptance.