How I verify legal claims

How this site is produced

A large part of the content here is produced by working with agentic AI. I am saying so up front, because this site rests on a claim about verifiability and concealing my own method would contradict it.

A second thing goes with it. None of this is written in one pass from an outline. I assemble it from my own attempts, notes, ideas and collected papers, and those hold other people’s thinking too, not only mine. Some of it came out of work on a specific matter, some as a note in the margin of something else, some is read and thought through after someone else. I put it together as I go, and where the origin of an idea can be traced, I name it.

That does not mean I have text generated and publish it. The agent works against a set of rules matching this methodology, and those rules are machine-enforced rather than merely recommended. In practice this is Claude Code with my own rule set and a library of legislation held on disk.

• Whenever a specific section or article is cited, the wording has to be opened in that library and checked word for word. The library holds texts from the Czech Collection of Laws and from the EU Publications Office database, not second-hand copies.
• For text that interprets a provision, getting the address right is not enough. Every interpreted article is read next to its wording, against four failure patterns I have been caught by myself: a rule imported from another instrument, a dropped word that changes how binding a rule is, a reference to a thematically adjacent place instead of the normative one, and a case-law summary written from the facts rather than from the ruling.
• Checks run before publication comparing the library against the sources and reporting mismatches. If they fail, nothing is published.

Review is ongoing, not finished. The rules and checks reduce the number of errors; they do not make the text error-free.

I also go over all of it myself as I go, and while doing so I add to it, correct it and improve it. When an error is found I log it in the list of changes, including what was wrong. Corrections of my own errors are there on purpose. That is why the list exists and why this site keeps growing and changing.

Sources

I take the wording of statutes from primary sources: the Collection of Laws (Sbírka zákonů) for Czech law, and the Official Journal and EUR-Lex for EU law.

Secondary sources, such as commentaries, scholarly articles, legal portals, or case-law digests, I use mainly as a guide. They help me find the connections, but they do not replace the text of the statute itself.

For my own work I also check citations against a local legal database before they appear in the output.

The three layers of a claim

In legal analysis I keep three things apart:

• Verified law The verbatim text of a provision in force, with a citation you can check against the primary source.
• Interpretation My inference, conclusion, or working opinion. It may be useful, but it is not the text of the law, so I do not label it as such.
• Missing facts Facts of the specific matter that I do not know and that could change the outcome. Rather than guess, I state them as questions or caveats.

The same logic runs through the prompt builder. Two notes go into it in more detail: Working with a statute and Why AI hallucinates on Czech law.

Citing sections

I never cite a specific section from memory alone. First I verify its current wording in the source, and only then do I work with it.

If I have no way to verify a provision, I say so plainly. I do not present the section number or its content as certain until it has been checked.

Effective dates and amendments

Statutes change. The same section can read differently in different periods, and older versions often keep circulating in articles, databases, and the output of language models.

So for legal claims I track not only the number of the provision but also the wording in force at the relevant date. Where it matters, I state the position as of a specific day.

Case law

Links to court decisions point to primary sources where possible, for example Curia, HUDOC, NALUS, or the websites of the Nejvyšší soud (Supreme Court) and the Nejvyšší správní soud (Supreme Administrative Court).

The annotations of decisions are my summaries. They are not a substitute for the full reasoning. Case law supports an interpretation, but before you rely on a decision in a specific matter you need to read it in full.

Revisions

I update the content as I go. Statute cards, research notes, and the course all carry a revision date, so you can see what state they refer to.

When a working hypothesis turns out to be wrong, I do not erase it without a trace. I mark it as disproven, so it is clear what changed and why.

For legislative processes that are not yet finished, I put the caveat directly in the text.

What this site is not

This site does not provide legal services, and its content is not legal advice for a specific case. It is research and educational material that shows the sources and the verification process.

Before you decide anything in your own matter, always check the wording in the source or consult a lawyer. Responsibility for using the information rests with the reader.

The research map shows how the individual working hypotheses come about.