Every intersection of Judaism
and computer science

Ten places where the two fields genuinely meet — not by metaphor, but because both are solving the same problem with different vocabulary. Each section below gives the Jewish source material first, then its computational mirror, then an honest accounting of what has actually been built here and what remains an open direction.

A note on method. It is easy to write that the Talmud "is like" a database and leave it there. That claim costs nothing and proves nothing. The sections below try to hold a higher bar: where a formal correspondence has been published in the peer-reviewed literature, it is cited as such; where a working system exists, it is linked; and where the connection is suggestive but unbuilt, it is labelled an open direction rather than dressed up as a result.

1. Formal logic and the thirteen middot

The source. The baraita of Rabbi Yishmael opens the daily liturgy with thirteen מִדּוֹת — measures, or rules — by which Torah is expounded. They are not rhetorical flourishes. They are named, enumerable operations: kal vachomer (קַל וָחֹמֶר, a fortiori), gezerah shavah (inference from shared wording), binyan av (generalization from a paradigm case), klal uprat (restriction of a general term by a following particular), and nine more. A student is expected to know which rule is being invoked, and to be able to attack an argument by showing the rule was misapplied.

The mirror. That is an inference system: a fixed set of typed rules over a body of text, with an explicit notion of validity. Computer science recognizes the shape immediately — it is a proof calculus. The interesting part is that the middot are not classically monotonic. Kal vachomer can be defeated by the objection dayo ("it is sufficient" — the derived case gets no more stringency than its source), which is a built-in limiting principle on how far an inference may carry. Rules that can be blocked by other rules are exactly what non-monotonic logic studies.

What is built. Dov Gabbay and his collaborators formalized this over roughly a decade in the Talmudic Logic series — kal vachomer reconstructed as matrix abduction, alongside deontic, temporal, and non-monotonic treatments of Talmudic reasoning. It is published mathematics, not analogy, and it is the load-bearing citation for most of this site.

2. Argumentation theory and machloket

The source. מַחֲלוֹקֶת — dispute — is not a failure state in rabbinic literature. It is the primary unit of content. The Mishnah routinely records the losing opinion by name, and the Talmud asks why: if the law does not follow Rabbi Eliezer, why preserve what Rabbi Eliezer said? The tradition's own answer is that a rejected position may be needed later, that a future court may rely on it, and — in the formulation that became famous — אֵלּוּ וָאֵלּוּ דִּבְרֵי אֱלֹהִים חַיִּים, "these and those are the words of the living God," even where only one may be acted upon.

The mirror. Modern argumentation theory models reasoning as a graph of arguments with attack relations between them, and asks which sets of arguments can coherently stand together. Dung's abstract argumentation frameworks and their many extensions — preference-based, value-based, bipolar — are the standard machinery. A sugya is strikingly close to an instance: positions, attacks (kushya), defenses (terutz), distinctions that dissolve an attack rather than defeat it, and the terminal state תֵּיקוּ — an unresolved stand-off that is recorded as such rather than forced to a decision.

What is honest. The correspondence is real but partial. Dung frameworks are abstract: they take the attack relation as given. A sugya's hard part is precisely deciding whether one statement attacks another, which is a question of textual interpretation, not graph semantics. Automating that is an open problem, not a solved one.

3. Knowledge graphs and the commentary tradition

The source. Open a page of Vilna Talmud. The Gemara sits in the center; Rashi runs down the inner margin; Tosafot down the outer; and around the edges sit cross-references to the codes, to parallel passages, to textual variants. The layout is not decoration. It is a rendering of a citation network, typeset in the sixteenth century, designed so a reader can traverse edges without leaving the node.

The mirror. That is a knowledge graph with a viewer. Nodes are passages; edges are typed relations — comments on, cites, explains, disputes, parallels. The modern equivalents are labelled property graphs and RDF, with the same central design question: what are the edge types, and who is allowed to add one?

What is built. A working graph of 157,143 nodes and 215,426 relationships across a 233-label schema. Its core is 17,147 rabbinic teachings, every one carrying a citable Sefaria reference, its Hebrew source text, and a 768-dimension vector embedding — 100% coverage on all three — drawn from 108 distinct source works and linked to 33,323 verses.

What is honest. 2,368 distinct relationship types over 215,426 edges is schema sprawl from uncontrolled extraction, not a hand-designed ontology. The graph is useful because it is cited, not because it is clean.

4. Version control and the Oral Law

The source. The Oral Law was transmitted in layers, each of which preserves the one beneath it. Mishnah (c. 200 CE) fixes a terse legal text. Gemara (c. 500 CE) layers argument onto that text without altering it. The Geonim, then Rambam's Mishneh Torah, then the Tur, then the Shulchan Aruch, then its glosses restate the law for new conditions — and each restatement is read against its predecessors, not in place of them. Pirkei Avot opens by naming the chain itself: Moses received the Torah at Sinai and transmitted it to Joshua, Joshua to the elders, the elders to the prophets.

The mirror. Immutable layers, additive commits, no destructive edits, and an explicit provenance chain from every statement back to a named source. That is version control with a blame view — and, unusually, one where the history is considered more authoritative than the tip. The transmission chain is machine-representable: sages as nodes, teacher-of and student-of as directed edges.

What is built. The extracted transmission network holds 2,504 rabbi-to-rabbi edges (1,191 TEACHER_OF and 1,313 STUDENT_OF) across 2,185 participating sages, whose largest connected component is 1,163 sages — 53.2% of the participants. A 150-node export of that component ships with the site.

What is honest. This is a machine-extracted network, not a historically validated one. It contains chronological errors — an edge from Moses to Avraham exists in the graph and is backwards by centuries — and transliteration aliases are unmerged, so a single sage can appear under six spellings and look less central than he was.

5. Gematria and number theory

The source. Hebrew letters carry numeric values, and rabbinic literature has used that fact as an interpretive device for two thousand years — sometimes as mnemonic, sometimes as homiletic support for a reading arrived at on other grounds, occasionally as an argument in its own right. The classical sources are notably self-aware about this: gematria is generally treated as asmachta, a support for a conclusion, not a proof of one.

The mirror. Computationally, gematria is a total function from strings to integers, which makes it trivially searchable and therefore trivially abusable. Over any large corpus, coincidences are guaranteed: the birthday problem does not care about theology. The honest computational question is not "do equal values occur" — they must — but "does a given correspondence occur more often than chance predicts in a properly specified null model," which is a hypothesis-testing problem with a well-known history of being done badly.

What is built. Tools that compute, index, and graph letter-value relationships across texts, so a correspondence can at least be checked rather than asserted. They are exploratory instruments, and no statistical significance claim is made for any pattern they surface.

6. Halakha as algorithm

The source. A posek deciding a case does not consult intuition. He runs a procedure. Is the case certain or doubtful (סָפֵק)? If doubtful, is the doubt in a biblical or rabbinic matter — because the default resolution differs. Is there a majority (רֹב) to follow, and is it a majority of a fixed set or an indeterminate one? Is there a presumptive prior state (חֲזָקָה) that persists absent evidence of change? Is the item nullified in a mixture, and at what ratio? Each of these is a named rule with conditions of application and known interactions with the others.

The mirror. This is a decision procedure over uncertainty: priors (chazakah), aggregation rules (rov), default logic for the underdetermined case, and an explicit meta-rule about which defaults apply in which domain. The structure is closer to a defeasible expert system with a documented conflict-resolution strategy than to anything probabilistic in the modern sense — though the rov literature's treatment of "follow the majority" versus kavua (a fixed item, treated as fifty-fifty regardless of population frequency) is a genuinely subtle statement about reference classes that would not embarrass a decision theorist.

What is honest. This is the intersection with the largest gap between what is describable and what is built. Formalizing a fragment of halakhic decision procedure as executable code is tractable; formalizing it in a way a posek would recognize as faithful is a research program, and nothing here claims to have completed it. Open direction.

7. Grounded AI and the citation imperative

The source. בְּשֵׁם אָמְרוֹ — "whoever reports a teaching in the name of the one who said it brings redemption to the world" (Pirkei Avot 6:6). Attribution is not a scholarly nicety in this tradition; it is a religious obligation, and the Talmud tracks chains of transmission with the seriousness of a provenance system. The corollary is equally strong: a teaching whose source cannot be named is suspect, and a scholar who does not know is expected to say so.

The mirror. Retrieval-grounded generation, verbatim citation checking, and abstention. A language model that produces fluent text without a retrievable source is doing exactly what the tradition forbids — reporting a teaching with no name attached. The engineering translation is a system that answers only from retrieved, citable passages, and refuses when retrieval comes back empty.

What is built. Eight conversational sages served over the cited Torah graph, with every claim checked against the source it names, measured under a zero-tolerance gate for fabrication on the grounded path. The live champion scores 9.25/10 grounding and 9.42/10 accuracy with citation-correctness 1.0 on a 12-probe battery — and the same system still slips on adversarial fabrication baits, which is documented rather than hidden.

8. Hermeneutics and interpretability

The source. פַּרְדֵּ"ס — PaRDeS — names four levels at which a text is read: peshat, the plain contextual sense; remez, the hinted or allegorical; drash, the homiletic and legal expansion; sod, the esoteric. The crucial claim is not that four levels exist but that they coexist without displacing one another. A drash reading does not overwrite the peshat; both are held, and confusing them is an error.

The mirror. Interpretability research keeps rediscovering this shape: the same computation admits descriptions at different levels — circuits, features, concepts, behaviors — and a higher-level story does not invalidate the lower-level mechanism. Marr's three levels of analysis make the same move. The rabbinic contribution is the discipline around it: a strict rule that a higher reading may not be used to contradict the plain sense in matters of law, which is a form of grounding constraint on abstraction.

What is honest. This one is a conceptual parallel, offered as a frame rather than a result. No system here implements PaRDeS as a representational hierarchy. Open direction, and an appealing one.

9. Calendar and computational astronomy

The source. The Hebrew calendar is a lunisolar calendar fixed by computation rather than observation, using a mean lunation of 29 days, 12 hours, and 793 chalakim (1080ths of an hour) — a value accurate to within about half a second of the modern figure — a nineteen-year intercalation cycle, and a set of postponement rules (dechiyot) that shift Rosh Hashanah off certain weekdays. Rambam devoted a section of the Mishneh Torah, Hilchot Kiddush HaChodesh, to the arithmetic, including spherical-astronomy procedures for predicting lunar visibility.

The mirror. That section is, in substance, an algorithm specification: integer arithmetic, modular cycles, conditional postponements, and a worked visibility computation — written in the twelfth century for readers who would execute it by hand. It is one of the cleanest examples anywhere of a legal text that is also a program, and it has the property good specifications have: two competent implementers get the same answer.

What is honest. The calendar arithmetic is thoroughly implemented elsewhere by others and well documented in the computational-calendrics literature. It is included here because it is the strongest single counterexample to the claim that the Judaism/CS connection is merely metaphorical — but it is not something this site rebuilt.

10. The epistemics of disagreement

The source. Beyond preserving disputes, the tradition theorizes them. Pirkei Avot distinguishes a dispute "for the sake of Heaven," which endures, from one that is not, which does not. The Talmud asks how the law can follow Beit Hillel when Beit Shammai were sharper, and answers with a procedural virtue: Beit Hillel taught their opponents' view before their own. Elsewhere a heavenly voice endorsing one side is explicitly overruled — לֹא בַשָּׁמַיִם הִיא, "it is not in heaven" — because the decision procedure, once given, belongs to the court.

The mirror. Every one of those moves is a live question in the epistemics of disagreement and in the design of systems that must act under unresolved dissent: when does disagreement license suspension of judgment, how should peer opinion be weighted, and what happens when an oracle of higher authority conflicts with an established procedure. The rabbinic answer to the last one — the procedure wins — is a remarkably strong statement about the value of auditable process over unauditable authority, and it is the single idea from this material with the most direct bearing on how AI systems should be governed.

What is built. The Sugya Protocol is the attempt to state that discipline in engineering terms: preserve the dissent, label the source, never delete, and let reinterpretation rather than erasure carry the system forward.

Where to go next

Every figure on this page is traced to a source ledger and re-verified before publication. Where a section says "open direction," nothing has been built yet and no result is being claimed.