In one sentence
We run two knowledge networks — the law as a citation network and every case file as a fact network — and in this sprint we connected them to each other for the first time. What is also new: you can finally see the first of the two as a network.
The first network: the law
The Rechtsbibliothek has been live since July 2026, and it is not a text archive but a citation graph: 424,855 decisions and 7.4 million citation edges. Every decision knows which statutes and rulings it cites — and by whom it is cited in turn. Clicking from a provision to a decision that cites it, and on to that decision’s own references, is a movement within the network, not a new search.
What is remarkable here is which of the two windows into a case-law database is the wider one. We hold 20,239 decisions in full text — which means a full-text search across case law is structurally blind to the vast majority of the collection, and most search masks do not say so. The citation graph, by contrast, knows all 424,855. Anyone who goes by citations sees more than anyone who goes by wording. How the library came about and where the data comes from is set out in the launch post.
What has changed in the library
It was public — but effectively invisible. That has been fixed:
- A network view on every detail page. A toggle between List and Relationships shows the citation neighborhood of a node as a graph. The list stays the default: it is the better form for reading up, the network the better one for understanding the surroundings.
- Every node has a page title of its own. Previously around 95,000 deep pages carried the same generic title — equally worthless for search engines and for your own browser tab.
- § cross-references in statutory text now jump instead of reloading the page. When reading a statute you want to read on, not start over.
- Mobile navigation exists at all. Below 1024 pixels, “Laws”, “Decisions”, “Legal concepts” and “Changes” were unreachable until now.
- A search result can be shared. Search term, court, year and court level are now part of the address — the back button works, and a research result can be passed on to a colleague as a link.
The second network: your case file
Alongside it, since June, stands the network that the case file itself is: 14 fact categories as nodes, 30 controlled legal relationship types as edges — “contradicts”, “witness for”, “evidence against” and so on. The decisive difference from a mere keyword search lies not in the volume but in the evidence: every fact and every edge carries a quotation with a jump into the source document. We described the thinking behind it when introducing the fact graph; how to query it directly is covered in the post on the fact graph chat.
The seam — and why the right provision decides everything
Until now the two networks stood side by side without knowing about each other. New in the “Facts & Relationships” tab of the case file is the “Legal basis & case law” panel. It asks two things at once: which provisions belong to the area of law of this case file? And which belong to the fact and relationship types that actually occur in this case file — witness details or deadlines, for instance? For those provisions it shows decisions that cite them. This is a pure database route: no AI call, no tokens, no credits.
While building it, a measurement taught us a lesson no textbook offers. The obvious choice for the witness axis would be § 373 ZPO — the provision on witness evidence. Except that courts barely cite it in their grounds of decision: together with § 1 KSchG it appears in 16 decisions, together with § 17 StVG in 4. What courts really write about when weighing a witness statement is § 286 ZPO, the free evaluation of evidence — the same intersections yield 265 and 446 there. We therefore curate the mapping according to what courts cite, not according to what would be systematically applicable. And every mapping is measured as it is created: anything that fails to reach a minimum number of shared references does not enter the collection. A class of pointers that reliably finds nothing is worse than none at all.
So that you do not have to read up on a hit in a second tab, a click opens a preview right inside the portal — the case file stays where it is.
What is not in here
A knowledge network is only worth something once every edge can be read up. So the limits belong in the same text, not in the small print:
- Sorting is by citation frequency — which does not mean “still good law”. Whether a decision has been overturned or superseded is something we do not show, because our source data does not carry that type of edge. This sentence appears in the panel itself, too, not just here.
- There are two edge types, no more: cites and is cited by.
- The link between fact and provision is a research basis, not an established fit. What is proven is that a decision cites the provision — not that it fits your case. That is why the interface never says “matching rulings”, but “Decisions citing § X”.
- “Nothing found” expressly does not mean “there is nothing”. Our library is incomplete, and only a portion of German case law is published at all.
- We hold no licensed commentary literature. Anyone who needs a commentary passage still looks it up where it is licensed.
Why we nevertheless consider this the right path
The last point is not an oversight but a positioning. We do not win on “more content” — but on the fact that every edge can be read up. That is precisely what decides whether a legal AI result is usable: a fabricated reference looks deceptively like a real one for as long as nobody checks. What happens when that goes wrong is shown by the Berlin case of invented rulings; what a firm can do about it is set out in the post on hallucinations in the law firm.
Both networks can be viewed without an account, or in the portal alongside your case file: the library at Rechtsbibliothek, the overview for law firms at Citation network for law firms and The case file in the lawyer portal.
