Case study

The hardest citation problem I could find

What makes a citation hard? Not finding a source - finding the right one, and showing the reader where it lives so they can check it themselves. Where is that harder than in a library argued over for two thousand years?

The main page tells the career. This page is the proving ground: a classical Hebrew library - thousands of interlinked texts, chains of teachers citing teachers across centuries - turned into a knowledge graph where every answer arrives with a citation you can look up and check. It began as a personal practice: close reading before work for years, and an archive I started saving in 2013. Everything below is what that practice became when the engineering caught up with it.

Dates, since a case study should carry them: the archive begins in 2013, the cluster it runs on has been in the house since 2023, and the graph, the citation verifier and the evaluation suites were built and put to work in 2026.

  1. Act I · The problem

  2. Why this corpus

    The benchmark

    A library built out of citations

    Where do you prove a system that must never guess? On the corpus with the strictest referees you can find. Classical Jewish texts are a citation graph by construction: every commentary names its sources, every source names its verses, and the chain of transmission is part of the text itself. Readers check references as a reflex, and precision is the whole culture. If a grounding system can satisfy this library's standards, easier corpora follow. That is why it is my benchmark.

  3. The data

    2013 → now

    Thirteen years of collecting

    In 2013 I started saving classes, essays and teachings to disk - before I knew what the collection was for. Thirteen years later it is the training and evaluation corpus: real teachings, each one traceable to the Hebrew source it turns on.

  4. A note I wrote myself, one month before my first day in blockchain

    “These books look like any other. But they are actually the records of our sages, generation after generation on the same source, the Torah. It's like the blockchain of history, it's been copied exactly, generation to generation, with very specific rules.”

    I then spent nearly ten years - ConsenSys, Chainlink, and a rack of GPUs at home - building the technology that could test that idea.

  5. Act II · The system

  7. Engineering

    The graph

    Every teaching carries its source

    The archive became a Neo4j knowledge graph: teachings linked to the Hebrew passages they interpret, authors joined to the teachers who taught them, verses connected to every commentary that turns on them. Ask a question and retrieval must land on a node with a real source attached - the graph is the ground truth the voices stand on.

    The graph today: 157,143 pieces - 17,147 teachings, all sourced; 33,323 Bible lines they join to; a teacher-to-student chain of 2,185 names. Counted from the live database on 2 Aug 2026; every query is published at jew.tech/evidence.

  8. Engineering

    The verifier

    Every citation, checked

    A verifier checks every citation against the original Hebrew before anything ships: in written evaluation suites, and again in live conversation. Only a model that clears every check goes live - alignment enforced as a release gate rather than a prompt, so a model that starts inventing sources cannot reach a reader. The methodology, the queries, and every number are published so anyone can re-run them.

  9. Engineering

    The voices

    Eight voices, out loud, from my house

    On top of the graph: eight speaking personas - open-weight models I can open up and inspect, served with Ollama across the GPU cluster in my house, with retrieval over the graph and speech generated locally. Data never leaves the building. Anyone can talk to them at app.openrabbi.com. It is grounded AI running in public, on the corpus with the strictest referees I know.

  10. Act III · Measured, in public

  11. Now

    Check my work

    You do not have to take my word for any of this: the current pass rates, the query sets, and every figure are published where anyone can re-run them. The research is at jew.tech: the evidence at jew.tech/evidence, the counting at jew.tech/methodology, the paper at jew.tech/sugya-protocol/paper. The productized form is Digital Twin Pro.

    If your team has a corpus where wrong answers are expensive - law, medicine, finance, safety - this discipline transfers. Write to me - I will answer, with sources!