LEGAL SOURCES THROUGH TIME

When a Correct CitationLeads to the Wrong Law

An artificial intelligence system can cite a genuine provision, a valid link and an official legal source, yet still give the wrong answer.

READING KEY

Temporal applicability
Legal RAG
Transitional provisions
Provenance
Official publication
Fail-closed control

Abstract. An artificial intelligence system may cite a real provision, provide a working link and point to an official legal source, yet still give the wrong answer. This happens when it relies on a version of the law that did not apply to the event under consideration. Provenance without temporal and jurisdictional applicability is incomplete provenance. In addition to the source itself, a legal AI system must preserve the specific version of the text, its relationship to the official publication, the temporal limits of its application, the relevant transitional rules and the path by which the system reached its answer.

On screen, everything looks as a lawyer might expect from a well-designed system. The answer is measured and precisely worded. It cites a statutory provision. The link works. The text of the provision is genuine.

There is only one problem: the question concerns an event in 2024, but the system has used the text as it applies in 2026.

This is not a conventional hallucination. The source exists, the provision exists and the quotation is accurate. What is wrong is its placement in time. That is precisely why this failure can be more dangerous than an obviously fabricated authority. The citation reassures the reader at the very point where it should raise a warning.

The central answer is straightforward: citing a genuine legal source does not by itself establish that the legal answer is correct. The system must also show which version of the source it used, in which jurisdiction, for which period and under which rules it concluded that the version applied to the facts in question.

A citation answers where the source came from, not when it applied

Reviewing a legal answer requires at least three distinct inquiries.

Check	Question
Authenticity of the source	Is this an officially published or otherwise authoritative text?
Identity of the version	Which specific temporal version of the text did the system use?
Temporal and substantive applicability	Did that version apply to the facts under consideration?

These questions are related, but they are not interchangeable. An authentic document may be the wrong version. A correctly identified version may belong to a different period or jurisdiction. A text may also have entered into force without yet being applicable.

A file hash can show that the bytes of a particular document have not changed since capture. It cannot show that the document was legally relevant. The European Legislation Identifier (ELI) can improve the identification and exchange of legislative data, while the European Case Law Identifier (ECLI) can identify judicial decisions unambiguously. Neither identifier decides, by itself, which law applied to a particular real-world event.

The technical integrity of a document and its legal applicability are separate evidential questions.

Law runs on more than one clock

In ordinary conversation, we often say that a law has “entered into force”. In legal analysis, that is rarely enough.

For a legislative act, the relevant dates may include at least:

the date of adoption;
the date of promulgation;
the date of publication;
entry into force;
the date of application of the act as a whole or of an individual provision;
the period during which a provision produces legal effects;
the date of an amendment, corrigendum, repeal or expiry;
transitional rules for pending proceedings and existing legal relationships;
the date of the event being assessed;
the date as at which the legal analysis must be made.

Article 154 of the Constitution of the Republic of Slovenia requires regulations to be published before they enter into force. Unless a regulation provides otherwise, it enters into force on the fifteenth day after publication. Entry into force, however, does not necessarily mean that every provision has also become applicable.

A useful and time-sensitive example is the Slovenian ZUOPP-1B, an amending act concerning the placement of children with special needs. This English description is unofficial; ZUOPP-1B remains the original statutory abbreviation. The Act was published on 17 February 2026, and Article 15 provides that it enters into force on the fifteenth day following publication. Article 14, however, provides that it becomes applicable only on 1 September 2026, subject to a special regime for one provision. Until then, the previous regime generally continues to apply.

On this article’s cut-off date, 26 August 2026, ZUOPP-1B was therefore already in force, but its new regime was not yet generally applicable. A system instructed merely to select the “latest law in force” could apply the wrong law to a case arising on 25 August 2026.

The same problem exists in EU law. Article 297 of the Treaty on the Functioning of the European Union links the publication of legislative acts in the Official Journal of the European Union to their entry into force. Individual acts, however, frequently specify a later date of application for the act as a whole or for particular obligations.

Legal time is therefore not a single date field. It is a relationship between multiple events and rules.

Applicable law cannot be found by sorting dates alone

Even a complete timeline of amendments is not yet a complete legal analysis.

Article 155 of the Slovenian Constitution generally prohibits the retroactive effect of legal acts. Exceptionally, only a law may provide that certain of its provisions have retroactive effect, where this is required in the public interest and no acquired rights are infringed.

In Decision U-I-13/13 of 14 November 2013, ECLI:SI:USRS:2013:U.I.13.13, particularly paragraphs 8 and 9 of the reasoning, the Constitutional Court of the Republic of Slovenia distinguished genuine retroactive effect from a new regime affecting only the future effects of an existing legal situation. The latter is not necessarily retroactivity within the meaning of Article 155, although it may engage the protection of trust in the law under Article 2 of the Constitution.

EU law draws a comparable distinction. In Land Nordrhein-Westfalen v Pokrzeptowicz-Meyer, C-162/00, 29 January 2002, ECLI:EU:C:2002:57, the Court of Justice explained in paragraphs 49 and 50 that new substantive rules generally do not apply to situations existing before their entry into force unless their terms, objectives or general scheme clearly require that result. New rules may, however, apply immediately to the future effects of situations that arose under the earlier rules.

For an AI system, the consequence matters. It is not enough to establish:

The event occurred on 15 March, and the new provision entered into force on 1 March.

The system must also examine:

whether the provision was already applicable at that time;
whether it governed the type of legal situation in question;
whether the legally relevant date is the date of the event, the filing of an application, the commencement of proceedings or the decision;
whether a transitional provision preserves the application of the previous law;
whether the case concerns a completed situation or the future effects of a continuing relationship;
whether a special regime applies to a particular provision, category of persons or type of proceeding.

These are not merely metadata fields. They are legal rules that the system must first retrieve correctly and then interpret.

A consolidated text is invaluable, but it is not always the final authority

Consolidated texts are almost indispensable in day-to-day legal work. They allow the reader to see an act with its amendments already incorporated, without manually assembling the original text and every subsequent amending act.

The redesigned PISRS is therefore an important infrastructural improvement. When launching it, the Office of the Government of the Republic of Slovenia for Legislation specifically highlighted text comparison, links between legal acts, an API and a digitised collection of unofficial consolidated texts. The word “unofficial” marks the evidential boundary. PISRS improves access to, comparison of and reuse of legal data, but where necessary the legal evidential trail must still reach the official publications, amending acts, corrigenda and transitional provisions.

At EU level, access to legal information must likewise be distinguished from authentic publication. Council Regulation (EU) No 216/2013 provides that the electronic edition of the Official Journal of the European Union, published in accordance with the Regulation, is authentic and produces legal effects.

The practical consequence is clear. A system must not retain only the text found on a convenient web page. It should also preserve:

the type of document;
the official publication source;
its relationship to the original act and subsequent amendments;
whether the text is original, corrected or consolidated;
the temporal scope of the specific version;
links to the acts that produced the change.

A consolidated text may be the best reading interface. It is not necessarily, by itself, complete evidence of how a provision produced legal effects on a particular date.

An identifier does not replace legal judgment

The European Legislation Identifier provides uniform URIs and structured metadata to improve the discovery, exchange and reuse of legislation. ECLI performs a comparable identification function for case law.

Such identifiers are foundational to sound legal data infrastructure. They reduce the risk of confusing documents and allow more stable links between systems. Identification, however, is not the same as applicability.

ELI can help a system determine which legal act or expression of an act it is handling. It cannot decide, in place of legal analysis, which real-world event is decisive, how a transitional provision operates or whether the new regime governs the future consequences of an older legal relationship.

The same is true of a hash. A hash proves the identity of a sequence of bytes. It does not prove the correct jurisdiction, legal validity or temporal applicability.

Static RAG can fail before the model writes a word

Retrieval-augmented generation systems, or RAG systems, first retrieve relevant passages and then supply them to a language model. The foundational architecture was introduced in Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.

In a legal setting, the problem arises when the search index contains only the current consolidated version of an act. The path may look like this:

the user asks about an event in 2024;
the system identifies the relevant act and provision;
the retriever finds the current 2026 version of that provision;
the language model produces a clear answer;
the system cites the correct provision and a working link.

Every technical operation may have worked as designed. The legal answer is nevertheless wrong because the corpus did not preserve the historical version, or because the retrieval layer did not apply a temporal filter.

A language model is not a reliable substitute for a missing legal source. If the relevant version was not retrieved, the generator cannot be expected to reconstruct its text, transitional regime and legal effects on its own.

Temporal validation must therefore occur before semantic ranking, or at least form a mandatory part of it. A suitable path is:

determine the jurisdiction and legal question;
identify the legally relevant dates;
retrieve all candidate versions;
filter them by periods of validity and application;
retrieve the amending acts, corrigenda and transitional provisions;
semantically rank the temporally admissible sources;
validate the citations and the hierarchy of legal sources;
answer, express qualified uncertainty or refuse to answer.

The generative model should receive a temporally validated evidence package, not an arbitrary copy of the latest text.

Research demonstrates the problem, not a universal solution

The preprint Temporal Misgrounding in Legal RAG: A Versioned-Corpus Benchmark for French Tax Law examines this precise failure mode. Its authors created FiscalQA Pro, a deliberately selected stress-test set of temporally sensitive questions in French tax law.

Across 209 scored questions concerning 33 provisions of the French General Tax Code, RAG over a static corpus of current versions achieved 2.7% mean strict accuracy. Retrieval over a multi-version index achieved 98.3% in the end-to-end setting, while a condition supplied with the correct provision in advance reached 99.1%.

These figures vividly demonstrate the failure mode under examination. They must not be presented as a general performance measure for every legal RAG system.

The benchmark was deliberately constructed so that, for 208 of the 209 scored questions, the current version did not contain the historical value sought. Most questions concerned time-varying numerical values, rates and thresholds. The final retriever used for the 98.3% result has not been publicly released, although model responses and tools for rechecking the reported results are available. The FiscalQA Pro repository therefore permits inspection of the published outputs and reproduction of part of the experiment, but not a complete independent reproduction of the final production retriever.

As at 26 August 2026, arXiv contains only version v1, submitted on 10 August 2026. The record states that the paper was accepted at the ICML 2026 AI4Law workshop; it does not identify a later peer-reviewed journal publication. The study should therefore be used as strong evidence of a specific temporal failure mode, not as a final or universal evaluation of legal AI.

A second line of research addresses the problem architecturally. An Ontology-Driven Graph RAG for Legal Norms presents SAT-Graph RAG, which separates an abstract legal act from its temporally specific expressions and models legislative amendments as queryable events. The work appears in conference proceedings, but its evaluation remains qualitative and proof-of-concept based. It is not yet a large quantitative benchmark.

The common lesson is not that temporal correctness has been solved. It is that temporal applicability should not be left to an instruction given only to the language model.

Time must become part of the data model

A legal information system needs more than a collection of text chunks and a last-updated field.

Akoma Ntoso 1.0 already provides important building blocks. It distinguishes the original work from individual temporal expressions of a document, records lifecycle events and supports separate representation of the beginning and end of force and efficacy for individual fragments.

For practical legal RAG, the data model should preserve at least:

a stable identity for the legal act;
the identity of the specific version or expression;
the identity of each provision through amendments and renumbering;
the jurisdiction and hierarchy of the source;
the date and source of official publication;
the period in force;
the period of application or legal effect;
the legislative events that produced each change;
linked transitional, repealing and corrective provisions;
the hash of the specific captured document;
the capture date and the date through which the corpus was verified.

It is also useful to distinguish two technical views of time. The first records when a rule was legally in force or applicable. The second records when a particular version was actually available in the data system. Without the latter, it may later be impossible to reconstruct reliably what the system could have known when it produced its answer.

The temporal verifiability envelope

Selecting the correct source is not enough. The result must carry sufficient information for another person to inspect the entire path.

We can describe this as a temporal verifiability envelope. For each material legal claim, it should contain at least:

the legal question and relevant date or period;
the jurisdiction applied;
the official identifier and a direct link to the source;
the specific version of the text and its hash;
the date of publication, entry into force and date of application;
the linked amending acts, corrigenda and transitional provisions;
the retrieval and temporal filters used;
transformations performed, including OCR, normalisation and chunking;
the source-retrieval date and corpus cut-off date;
detected conflicts or missing data;
the status of human review.

In A Result Is Not a Method, we argued that a legal result should preserve its sources, transformations, limitations and review path. The temporal verifiability envelope adds one indispensable question: why should this source be treated as the law for the time under examination?

For OpenLegalCore, this is principally an architectural criterion, not a marketing claim about functionality already implemented. Open source code can make rules for version selection, filters, resolvers and error handling inspectable. It cannot supply missing historical data or correct a flawed legal interpretation. Openness makes the method open to inspection; correctness must still be demonstrated.

If temporal applicability cannot be established, the system should say so

Sometimes the decisive date is missing from the legal question. The corpus may not contain the historical version. A transitional regime may be ambiguous or distributed across several acts. Even official metadata can be incomplete.

In such a case, the system should not silently select the latest version and generate a confident answer. A fail-closed approach is more appropriate:

identify the missing information;
display the candidate versions;
explain the temporal conflict;
request the additional fact;
refer the question for human review.

A refusal is not a technical failure if it prevents an apparently well-supported answer based on the wrong law. The reason for the refusal should itself remain visible and verifiable.

Four dangerous shortcuts

“Always use the latest text”

The latest text is appropriate only where the question concerns the current regime and no transitional or deferred application is relevant. For historical events, the latest text may be precisely the wrong one.

“The consolidated text is enough”

A consolidated text is an excellent reading tool, but it may obscure which amending act produced a change, when that change became applicable and whether the consolidation is labelled unofficial. Where necessary, the evidential trail must reach the official publications.

“A correct link proves a correct answer”

A link proves only where it leads. It does not prove that the system used the correct version, the provision as it was structured at the relevant time or the correct transitional regime.

“The model will resolve the temporal question itself”

A model can explain retrieved sources. It should not be the sole mechanism guessing which version is relevant. The temporal choice must be visible in the data, the query and the audit record.

Five questions for reviewing a legal answer

Before a legal AI result is used in serious decision-making, the reviewer should receive clear answers to five questions:

Which real-world event and which date are legally relevant?
Which jurisdiction and hierarchy of legal sources apply?
Which specific version of the provision was retrieved, and from which official publication does it derive?
Were entry into force, the date of application, transitional rules and possible future effects on older legal situations checked?
Can the retrieval path be reproduced, with the limitations, conflicts and status of human review made visible?

If the system cannot answer these questions, the citation is not yet evidence of the correct legal basis.

Limits of this analysis

Version-aware and temporally aware retrieval addresses an important class of errors, not every legal problem.

It does not guarantee that:

the facts were established correctly;
the correct legal act was selected;
the corpus is complete;
the provision was interpreted correctly;
every material conflicting authority was found;
the case law was weighted correctly;
questions of private international law or conflicts between legal rules were properly resolved;
the human review was adequate.

This article focuses primarily on temporal versions of legislation and the technical retrieval of sources. It does not provide an exhaustive method for evaluating the temporal effects of case law, procedural rules or every category of continuing legal relationship. In an individual case, determining the applicable law remains a legal task.

A correct citation is where the proof begins

The wrong law does not always arrive as a fabricated provision. Sometimes it comes with an official title, a stable link and entirely genuine text.

Only one answer is missing: whether that text was the law applicable at the relevant time.

A legal AI system must therefore establish more than the origin of its answer. It must also establish the temporal and jurisdictional path to the source it applied. The latest text is not necessarily the applicable law, and a technically unchanged document is not necessarily a legally relevant document.

A citation without time can show where an answer came from. It cannot, by itself, establish that the answer came from the law that applied.

Selected sources

Legal and technical sources last verified: 26 August 2026.