The texts
The portal is meant as a growing resource. It starts with the early printed editions in order to make the corpus available quickly. An introduction explains what consilia are, how the collections came into being and why they have been difficult to access. Each author has a short page on his life, work, printed editions and the manuscripts of his opinions.
| Volume | Edition | Consilia | Source of the images |
|---|
How to use the portal
There are three main points of access to the consilia:
Read lets you study the consilia one by one in the traditional way, as given in the printed volumes, with an index on the left, the text in the middle and the digital facsimile on the right. Every consilium is reproduced as it appears in its source, often including the summaries that the prints place at its beginning. The facsimile follows the text, with boxes marking the columns of the consilium.
Search offers several analytical points of access to the text. Full text searches the Latin text, treating the usual spelling variants (i/j, u/v, ae/e) alike and finding all inflected forms of a word. Semantic search ranks passages by meaning for queries in Latin, English or German, using embedding models. Similar finds the passages closest to a chosen consilium, using the same models. Subjects browses machine-generated keywords and subject layers.
Ask offers an epistemic infrastructure based on external language models such as Claude, OpenAI models, Gemini or another compatible service, used through your own API key or endpoint. It allows complex analyses of the whole corpus in a transparent manner.
How the texts were made
This site rests on a human-in-the-loop pipeline that combines several deep learning models. The portal and the models are developed in the project Embedding the Past (LOEWE-Exploration, TU Darmstadt, AdWL|Mainz).
Transcription
The transcription of each print was made in a multi-layered pipeline. It starts with a layout analysis that recognises the text regions and annotates the lines by their structural function, checked by hand (data, doi:10.57967/hf/11321, and the layout model trained on it). Text recognition was first done with a PyLaia model in Transkribus; its output was corrected and used to train a kraken model. Because many domain-specific abbreviations were hard to transcribe consistently, the vision language model Gemini 3.8 Flash read every column a second time. Text on which at least two of the three readings agreed was adopted for this edition. Still, the transcription may contain errors. Abbreviations are expanded silently, long s is normalised to s, accents are disregarded, and lines are joined across hyphenation into the paragraphs of the print. The Panormitanus volume was transcribed by Gemini 3.8 Flash and then compared line by line with a second reading by the kraken models trained on Baldus. Where the two readings differ and only the kraken reading gives a word attested in the Baldus text, as with a long s misread as f, the text was corrected, and abbreviations left unexpanded were resolved.
Keywords
Keywords were extracted automatically by Qwen3.6-27B, clustered into a shared vocabulary with sentence embeddings (BGE-M3) and reduced to about ten terms per consilium. Gemini 3.8 Flash added some keywords in a second review run.
Subject layers
Two controlled subject layers were applied. The first uses a fixed vocabulary of legal subjects (such as succession, marriage and dowry, persons, obligations, procedure, criminal law, statutes), assigned per consilium as a central or secondary subject. The second, on issues such as body and sexuality, was annotated by Qwen3.8-27B twice, and only annotations found in both runs are included in the dataset. It is important to bear in mind that these annotations were made by a large language model: they offer solid access to the material but need to be used with some caution.
Citations and names
Legal citations and the names in the Latin text were annotated by two independent and complementary models. A sequence tagger (Flair, built on the MediBerta model, an XLM-RoBERTa continued on medieval Latin, together with two character language models trained on Latin legal text, forward and backward) was trained on annotated data and tagged the corpus. A second run was made with Gemini 4 Argon. Spans found by both models are shown as confirmed and spans found by only one of them as unconfirmed; citations found only by the tagger are left out. In general, citations of the civil law and jurists are the most reliable annotations, while places, institutions and dates show frequent errors. The tagger is published and can be reused. Citations are also resolved to a work and title ("ff. de pac. l. tale pactum" becomes Digest 2.14, De pactis) by rule: the siglum names the work, and the abbreviated words are matched as prefixes of the title names in tables compiled from the standard editions. Citations fitting more than one title stay unresolved. A bare lex or chapter that directly follows a resolved citation takes over its title, and for some consilia a title was added where two language models named the same one independently.
Search and the assistant
Semantic search is based on sentence embeddings at paragraph level, using BGE-M3, a BGE-M3 model fine-tuned on Latin legal text and Qwen3-Embedding. All vectors are published on Hugging Face as a dataset. Models can be selected on the search page, and queries are embedded in your browser unless you connect an API yourself. The full-text search finds all forms of a word through a lemma list made with the Latin lemmatizer of the project. A word in quotation marks is searched in the given form only. The Ask page lets a language model of your choice work with the same data through a set of tools that make the corpus accessible to the model. The model is prompted to cite its findings with the internal ids. The tools are also available to AI assistants through an MCP server, which can also perform semantic search with further models such as CanonBERTa, a Latin model fine-tuned on legal text.
Status and errors
The goal of this edition is to make a large and hitherto inaccessible collection accessible by means of machine learning. It will therefore contain errors in transcription, annotation and analysis, and should always be used with some caution and checked against the scan.
Citation and reuse
Cite a consilium with the reference the Cite button copies, for example:
The edition as a whole, including its workflows: Michael Schonhardt (ed.), Consilia. A portal for medieval legal opinions, data revision —, .
Data, code and TEI exports are available in the project repository. Text data are released under CC BY 4.0, the code under the MIT licence.
Imprint
Responsible for this site: Prof. Dr. Michael Schonhardt, Digitale Editorik und Kulturgeschichte des Mittelalters, Technische Universität Darmstadt, in cooperation with the Akademie der Wissenschaften und der Literatur | Mainz. Institut für Sprach- und Literaturwissenschaft, Residenzschloss 1, 64283 Darmstadt, michael.schonhardt@tu-darmstadt.de.
Corrections and suggestions: michael.schonhardt@tu-darmstadt.de. The portrait in the page head is Baldus de Ubaldis after the engraving in Jean-Jacques Boissard, Bibliotheca sive Thesaurus virtutis et gloriae (Frankfurt 1627), via Wikimedia Commons, public domain.