famulus-reference

A citable, local and trustworthy ground for AI in the humanities

The claim

An AI answer about our cultural heritage is only as good as the source it can point to. famulus-reference makes this the rule rather than the exception: every document, page and passage in a collection receives a stable address, a citation derived from recorded bibliographic data, and a documented chain of provenance. A model that works on such a collection cannot invent a reference. It can only quote one that exists, or say that the reference is missing.

The problem it solves

Today’s AI systems answer fluently and cite unreliably. Sources are invented, provenance is lost, and the material to be interpreted, often unpublished and often sensitive, has to be sent to servers outside the institution and outside Europe. For archives, museums, academies and universities this is neither scientifically nor legally acceptable.

The concept

famulus-reference treats every collection as a reference data source in need of famulus proof.

  • An address for everything. fr:<collection>/<document>/<page>/<segment>, persistent and used verbatim in every claim the system makes.
  • Citations that are never invented. BibTeX and citation text are generated only from recorded bibliographic data. Where data is missing, the gap is stated.
  • Graded evidence. Each statement carries a grade: documented, inferred, hypothesis or gap, together with the checks that produced it.
  • Full provenance. Which tool, which model, which instruction and which input produced which output is recorded, and every query leaves a trace.
  • One universal service. All collections live in one DuckLake and are served by one MCP server, so any AI assistant, notebook or search tool reaches them in the same way. The same data can be harvested by standard catalogues (OAI-PMH, JSON-LD/DCAT, IIIF) and federated between institutions without giving up control.

Local, confidential, European

famulus-reference is built to run on the institution’s own hardware.

  • Models and data stay local. Inference, indexing and storage run on our own machines (Mac Studio, private network). No document and no query is sent to a cloud service.
  • Confidential by design. The collection is written by exactly one machine and served read-only. Rights holder and licence are recorded per collection, and access is by explicit grant.
  • Traceable by design. Every answer can be reconstructed from its trace. This matches the transparency, documentation and data-minimisation duties of the GDPR and the EU AI Act. These are design properties of the system, not a certification.
  • Culturally autonomous. A collection can be shared, harvested or federated on its owner’s terms, and it can be removed again.

What it looks like in practice: the Kármán Nachlass

The papers of the aerodynamicist Theodore von Kármán (1881–1963) are our first large test case:

Documents 1,068
Pages 38,172
Letters, 1900–1963 5,676
Persons 11,233
Institutions 4,203
Places 2,872, in 31 countries
Languages of the letters English, German, Hungarian, French, Italian, Spanish

Such a Nachlass is normally searchable only by someone who knows what to look for. With famulus-reference the collection becomes a citable partner for questions:

  1. Find and cite. “Which letters discuss the wind tunnel at Aachen?” The answer lists the letters with page-exact addresses and ready-made citations.
  2. Follow people and places. Networks of correspondents, institutions and travels, each link resolving to the passage that justifies it.
  3. Compare across collections. The same questions, asked across Kármán, Kepler, Hartmann or Pirckheimer, with each statement staying attached to its own collection.
  4. Support editions and teaching. Transcriptions, translations and commentary that cite their sources by construction.
  5. Open the door to other holdings. Scientific estates, museum archives, university collections and regional archives enter through the same intake procedure, and keep their own data, rights and identity.

The Kármán figures are taken from the collection database (14 October 2026). The Hartmann Collectio figurarum is the pilot collection for the lake itself.

Why it matters

A knowledge society that relies on AI needs a ground that can be checked. We aim for famulus-reference to become that ground:

  • Trust, because every claim points to a source that can be looked up.
  • Epistemic reliability, because evidence is graded and gaps are admitted.
  • Cultural autonomy, because the knowledge of a community is held, run and governed by that community, and not delivered to a platform.

These are goals that the system has to earn through measurement and public audit. It is designed to make that earning possible.

What support makes possible

  • bringing further collections into the lake through the intake procedure,
  • peer nodes, so that institutions can federate their collections while keeping control,
  • systematic evaluation of reliability, published openly,
  • long-term operation and maintenance of the infrastructure.

For all papers and technical documents of the platform, grouped by aspect, see the Famulus landing.

For the technical architecture see the platform paper and the orchestration chapter.

Contact: to be added.