Famulus Platform

Famulus platform · Start here

Famulus.
AI that shows its sources.

Famulus is a way of working with AI in the humanities and sciences in which every answer can be traced back to a source. Models and data stay on the institution's own machines, every page and passage of a collection has a permanent address, and every result carries a record of how it came about.

Gerd Graßhoff, Humboldt-Universität zu Berlin, Forschungsinstitut des Deutschen Museums

Typeset papers are alpha drafts. Technical documents are living pages of the platform documentation. Nothing on this page is published beyond the drafts already released by the author.

12typeset papers
322pages in total
4documents
12edition landings

Famulus in brief

A famulus is, in the old university sense, the assistant who serves a scholar. The agents of this platform are meant to be exactly that: assistants who show their working.

The problem

AI models answer fluently and cite unreliably. They invent sources, forget where a statement came from, and, to work on unpublished or sensitive material, usually need it to be sent to servers outside the institution.

The idea

Treat a collection as a reference that has to prove itself. Each document, page and passage gets a stable address; each citation is composed only from recorded bibliographic data; each statement carries a grade of evidence. A model working on such a collection can quote a reference that exists, or say that it is missing.

What you get

Answers that can be checked line by line, work that can be reproduced from its trace, and an infrastructure that runs on your own hardware. Researchers work in notebooks, where text, code and sources stand side by side.

Three ways into this page

I am new hereRead the introduction to Famulus Reference first, then the platform paper.Start here →I work with sourcesSee Famulus at work on editions, letters, tablets and objects.Applications →I build or evaluate systemsSpecifications, score language, benchmarks, epistemology and drafts.For specialists →

Three ideas that carry the platform

Everything else on this page is an elaboration of these three.

First idea

Famulus Reference

A global reference for AI models

Each collection is served as a reference data source. A model asks it for grounding and receives not only the text but the address, the citation, the provenance and the grade of evidence of every item it may use.

What it achieves

  • An address for everything. The address is persistent and is used verbatim in every claim the system makes.
  • Citations never invented. A citation is composed only from recorded bibliography. Where data is missing, the gap is stated.
  • Graded evidence. Each statement is documented, inferred, a hypothesis, or a gap, together with the checks behind the grade.
  • A trace for every answer. Which tool, model and input produced which output is recorded; a later request can read the same items in the same state.

How it works

All collections live in one local lake of data and are served by one MCP server, so any AI assistant, notebook or search tool reaches them the same way. The same data can be harvested by library catalogues (OAI-PMH, JSON-LD/DCAT, IIIF). Any institution can run the same server over its own lake, so the reference is distributed rather than centralised; the registry that would connect such nodes exists as a stub, and the server's deployment on the central machine is prepared.

Read the introductionTechnical specificationThe Famulus universe

Second idea

From literate programming to literate AI

Knuth's root, and what Famulus makes of it

“Instead of imagining that our main task is to instruct a computer what to do, let us concentrate rather on explaining to human beings what we want a computer to do.”

Donald E. Knuth, Literate Programming, 1984

Knuth proposed that a program be written as a document for human readers, with the code woven into the explanation. Famulus takes this over for AI. The explanation becomes the specification an agent works from, and the sources it cites become addressable.

  1. 1984

    Literate programming

    Program and prose in one document, written for people. Knuth's root.

  2. The method

    Literate coding with AI agents

    Theory, data schemas, thresholds and code in one executable document, written before the agent starts.

    Read the paper ↓
  3. The agents

    AI Partitur

    The agent itself is a readable score: a YAML document that a runtime executes. Everything that decides validity is an open field.

    Read the paper ↓
  4. The medium

    Notebook and Reference

    Text, executable blocks, results and a verdict against an independent expectation. Every cited source resolves to an address.

    Read the paper ↓

Literate programming made programs readable. Literate AI makes AI work readable, checkable and citable.

Third idea

Certified corpora and arguments

Where the sources come from, and how an answer follows from them

PDAI, the corpus

Scanned editions become structured corpora. A page enters the corpus only after it has passed an audit, and the page is the unit of citation.

Read the paper ↓

Reasoning, the argument

A question to a collection becomes the conclusion of an argument. Facts are premises, critical questions search for objections, and a conclusion is validated only if it follows. An objection found once becomes a lesson for all arguments of its kind.

Read the paper ↓

Start here

7 texts in a reading order for newcomers. The first is written for a general reader; the others deepen one idea each.

1

FAMULUS REFERENCEReadable introduction

Famulus Reference: a citable, local and trustworthy ground for AI

Begin here. What the reference system promises, in plain language, with the Kármán papers as the first large test case.

Every document, page and passage of a collection receives a stable address, a citation derived only from recorded bibliography, and a documented chain of provenance. A model working on such a collection cannot invent a reference; it can quote one that exists, or say that it is missing.

Open the document

2

THE PLATFORMTypeset draft

Famulus. Eine verteilte lokale Plattform für professionelle Forschung

Read second. The whole architecture: who provides what, and how agents back every step with a checkable source.

Famulus as a distributed local platform. A language institution, a knowledge base and a methods institution carry agents that back every step with a checkable source. Researchers work in notebooks; models and data stay on the institution's own machines. (The paper is in German.)

25 pages · September 2026 · Gerd Graßhoff

PDF

3

LITERATE AITypeset draft

Literate Coding with AI Agents: A Methodology for Computational Analysis of Scientific Objects

Complexity Reduction through Structured Context, Executable Specification, and Closed-Loop Refinement

Read for the idea. How Knuth's literate programming becomes a method for working with AI agents.

A method for joining literate programming and AI agents. Before anything is handed to an agent, a single executable document is written that holds the theory, the data schemas, the thresholds and the code. The worked example is the analysis of conic sundials.

5 pages · September 2026 · Gerd Graßhoff

PDF

4

THE NOTEBOOKTypeset draft

Famulus Notebook

A teaching book: writing, running, checking, annotating, sharing and publishing scholarly notebooks

Read to see the work in practice. A textbook with six research projects.

The notebook as the working medium: text with executable blocks, followed by the result, its artefacts and a verdict against an independently determined expectation. Six research projects show the notebook at work.

114 pages · September 2026 · Gerd Graßhoff

PDF

5

CERTIFIED CORPORATypeset draft

PDAI: A Certified Document Infrastructure for Computational Historical Scholarship

From PDF to a Queryable, Audited, and Learning-Enabled Corpus

Read to learn where the sources come from. How a scanned edition becomes a corpus that can be cited page by page.

PDAI, a certified document infrastructure that turns scanned scholarly editions into structured, queryable corpora. A page enters the corpus only after the audit, and every error is assigned to one of four causal layers.

21 pages · September 2026 · Gerd Graßhoff

PDF

6

REASONINGTypeset draft

Famulus Reasoning

Evidence, reasoning chains and validation for questions to scholarly collections

Read to see how an answer is argued. From a question to a conclusion with addressed premises.

A question to a collection becomes the conclusion of an argument. Addressed facts of the collection database are premises, critical questions search for objections, and a conclusion is validated only when it follows from its premises. An objection against a reason becomes a checked lesson for all arguments of its kind.

16 pages · October 2026 · Gerd Graßhoff

PDF

7

IN USETypeset draft

AI Erasmus

The philosophy, the operations of an AI famulus for scholars

Read to see a research platform built on it. Editions as the substrate of inquiry.

A guide to AI Erasmus, a research platform that treats editions as the substrate of research: reading with an edition, querying with a visible method, dictionary, persons and correspondence, comparison of concepts.

23 pages · September 2026 · Gerd Graßhoff

PDF

Famulus at work

Famulus is applied on real collections. Each card opens the landing page of an edition or research project, with its own numbers, narrative and open points. Figures are read from the sources when the page is built.

Astronomy and the exact sciences of the early modern period

Collected works and letters

Johannes Kepler

The Collected Works as full text, a register of letters, and Famulus notebooks that put questions to the Latin text, such as lex and natura, or anima and vis.

26volumes
14,277pages
934letters
Open the landing →

Opera omnia

Tycho Brahe

The Opera omnia in Dreyer's edition, with the tables of the transcription and the star catalogues.

14volumes
5,864pages
10,164tables
Open the landing →

Works and literature

Nicolaus Copernicus

The works and editions of Copernicus and the secondary literature, with the languages of the corpus.

10primary works
3,644pages
56secondary works
Open the landing →

Correspondence

Johannes Hevelius

The four volumes of the Correspondance, with the state of the scans.

4volumes
2,254pages transcribed
Open the landing →

Opera omnia and letters

Leonhard Euler

The Opera omnia, the Eneström index and the correspondence of series IVA.

84volumes
41,219pages
866index entries
Open the landing →

Ancient and medieval sources

Astronomical Diaries

Babylonian astronomy

The Astronomical Diaries as a cuneiform tablet edition under its own DTD, with a layered audit and a famulus-reference extension.

7volumes
2,333tablets
86tablets edited
Open the landing →

Measured on 3D models

Ancient sundials

The Berlin register of sundials, measured on the 3D models. The pilot case of literate coding with AI agents.

782dialface entries
113measured and graded
72latitudes determined
Open the landing →

Bischoff's catalogue

Medieval manuscripts

Bischoff's catalogue of the continental manuscripts of the ninth century, with holdings, places and the structure of an entry.

3volumes
1,641pages
10,018entries
Open the landing →

Philosophy, physics and estates

Papers and letters

Theodore von Kármán

The papers of the aerodynamicist, with correspondence, manuscripts and Famulus notebooks. The first large test case of Famulus Reference.

Open the landing →

Regularity theory

Causality

The regularity theory of causal relevance by Graßhoff and May, its picture language, the case book and the program.

16documents
353pages
16tools
Open the landing →

The mechanical background

Wittgenstein and mechanics

Hertz's Principles of Mechanics, Reuleaux's Theoretical Kinematics and Graßhoff's papers on the mechanical background of the Tractatus.

6works
1,153pages
Open the landing →

Genesis 1918 to 1922

The Tractatus typescripts

The typescripts and corrected copies of Wittgenstein's Tractatus, with the history of their genesis.

473pages
351plates
Open the landing →

For specialists

Specifications, the score language, benchmarks, the epistemological foundation and unfinished drafts. These texts assume that you know the platform; the introduction above does not depend on them.

Grounding, reference and storage

SPECIFICATIONTechnical specification

famulus-reference: address, schema and API

The fr: address scheme, the schema famulus-reference/1, the MCP tools, run traces and snapshots, and the harvest routes (OAI-PMH, JSON-LD/DCAT, IIIF). States what is implemented and what is a stub.

Open the document

ARCHITECTURELayers and components

The Famulus universe

Every layer between a source folder and a published landing page, with its component, path, port and role. The map from import to evidence to landing, and the machine topology.

Open the document

OPERATIONSOperations and reference

Platform documentation

Installation, routes, services, collections and the notebook in operation, on the platform documentation site.

Open the document

Agents, verification and benchmarking

CRITICAL ANNOTATIONTypeset draft

The Famulus Probe Lab

Normalise, filter, classify: generating critical annotation for the items of a reference storage

How a probe, one sentence of the scholar, produces critical annotation for the items of a reference storage in three stages: normalise (every expression gets a footnote with grounds and a grade, no value without a ground), filter (the justified keys of the probe become one query) and classify (entailment, contradiction, independence, and weak or abductive support). One worked case on the Kármán Nachlass; the limits are stated.

19 pages · October 2026 · Gerd Graßhoff

PDF

ORCHESTRATIONTypeset draft

The AI Partitur

Orchestration one can read: agents as movements, YAML as notation

A declarative score language for agents. Each agent is a YAML document executed by a runtime (the conductor); every parameter that decides scientific validity is an open field, so the workflow can be read and corrected by the scholar.

12 pages · September 2026 · Gerd Graßhoff

PDF

VERIFICATIONTypeset draft

Golden Queries in AI Erasmus

A testbed for scholarly inquiry: query forms, grounding, and the monitor log

Thirteen canonical research questions with the answer declared in advance, checked by a counting oracle and a lower-bound oracle. The monitor log makes every answer inspectable.

22 pages · September 2026 · Gerd Graßhoff

PDF

COMMUNICATIONTypeset draft

AI-Reporter: A Path to a New Genre of Scientific Communication

From Presentation to Publication through Agentic LLMs

An overview of the history of scientific communication from manuscript to print, from Regiomontanus through Copernicus and Kepler to the Royal Society. The AI-Reporter continues this line and turns talks into assessable records within minutes.

20 pages · December 2025 · Gerd Graßhoff

PDF

Epistemology, language and the lexicon

LANGUAGETypeset draft

Wittgensteinian Grammar and OED Dictionaries for AI

A foundation of semantic understanding of scientific texts

Formulation audit against the entry form of the Oxford English Dictionary. Whether a formulation preserves meaning is decided by rule compliance anchored to cited pages.

25 pages · October 2026 · Gerd Graßhoff

PDF

EPISTEMOLOGYTypeset draft

Computational Epistemology: Foundations

Evidence and Its Justificatory Reach in the Sciences and the Humanities

The methodological foundation: how knowledge claims in historical science can be evaluated computationally, and why the evaluation must be traceable to recorded evidence. The source still carries the note "DRAFT SKELETON, NOT FOR CITATION".

20 pages · October 2026 · Gerd Graßhoff

PDF

Drafts and skeletons

Not typeset yet. Listed for completeness, without a link.

TitleStateWhat it is about
Querying with AI Linguistic TreesLaTeX draftA linguistic index layer for large historical collections: language detection per unit, Stanza analyses next to each document, and a DuckDB schema for lemma and predicate-argument queries.
PDAI: A Page-Centric Document Format and Processing Pipeline for AI-Driven Scholarly Archive ProcessingSkeleton, title and abstract onlyPDAI as a page-centric document format with an immutable page layer and a derived segment layer that keeps the provenance of the pages.
A Multi-Layer Index and Retrieval Architecture for Historical Document CollectionsSkeleton, title and abstract onlyA six-layer index architecture for historical collections: entity extraction, token index and a concept model in the sense of the Oxford English Dictionary.
AI Scholar: From PDF to Portable Document AI InfrastructureSkeleton, title and abstract onlyAn architecture of three packages (ai-scholar, ai-viewer, ai-classifypdf) for processing scholarly documents from PDF to structured PDAI output.
A Hierarchical Modular Skill SystemSkeleton, title and abstract onlyA hierarchical system of declarative skills that gives AI agents current knowledge of tools and workflows across the coordinated projects.

Open points

  1. Famulus Tabula Rudolphina is a sketch (a chat analysis of Kepler's star catalogue, 2026-10-08) and is not listed here until it has a finished draft.
  2. Papers are drafts; none is published with this page. Freigabe is pending.
  3. The English summaries of the platform paper and the other German-language drafts are condensed by hand from their text; the author should check them against the PDFs.
  4. The technical documents link to the platform documentation on the hub. They are not part of the paper archive.
  5. Famulus Reference is implemented as a server and a schema; the registry for peer institutions is an empty stub and cross-institution queries are not implemented.
  6. The Kármán landing has no figures here because its collector data is not part of this build.

Not listed as a paper

  1. famulus-tabula-rudolphina: Sketch; no finished paper yet.
  2. ai-voicememo-visionary-platform: Skeleton (title and abstract only).