At first, half its picks were wrong.
The first version, in January 2026, flagged legal citations (“Rabbi X says in the name of Rabbi Y”) as stories. The expert's first review showed us why. Each review since has made it better.
All projectsIn progress
מַעֲשֶׂה בְּ…
Ma’aseh b’… “It once happened that…”: the Talmud’s way of saying story time.
Finding the stories in the Talmud
The Babylonian Talmud is mostly legal argument, but stories are scattered through it: rabbis travelling, quarrelling, getting things wrong, being kind. No index tells you where they are. This project teaches an AI to find them, and checks its work against a scholar who has spent his career on exactly these stories.
Scroll, and watch two stories surface
Ketubot 67bכתובות סז:
תָּנוּ רַבָּנַן: ״דֵּי מַחְסוֹרוֹ״ — אַתָּה מְצֻוֶּוה עָלָיו לְפַרְנְסוֹ, וְאִי אַתָּה מְצֻוֶּוה עָלָיו לְעַשְּׁרוֹ. ״אֲשֶׁר יֶחְסַר לוֹ״ — אֲפִילּוּ סוּס לִרְכּוֹב עָלָיו וְעֶבֶד לָרוּץ לְפָנָיו. אָמְרוּ עָלָיו עַל הִלֵּל הַזָּקֵן שֶׁלָּקַח לְעָנִי בֶּן טוֹבִים אֶחָד סוּס לִרְכּוֹב עָלָיו וְעֶבֶד לָרוּץ לְפָנָיו. פַּעַם אַחַת לֹא מָצָא עֶבֶד לָרוּץ לְפָנָיו, וְרָץ לְפָנָיו שְׁלֹשָׁה מִילִין. תָּנוּ רַבָּנַן: מַעֲשֶׂה בְּאַנְשֵׁי גָּלִיל הָעֶלְיוֹן שֶׁלָּקְחוּ לְעָנִי בֶּן טוֹבִים אֶחָד מִצִּיפּוֹרִי לִיטְרָא בָּשָׂר בְּכׇל יוֹם. לִיטְרָא בָּשָׂר מַאי רְבוּתָא? אָמַר רַב הוּנָא: לִיטְרָא בָּשָׂר מִשֶּׁל עוֹפוֹת. וְאִיבָּעֵית אֵימָא בְּלִיטְרָא, בָּשָׂר מַמָּשׁ.
מאי קא משמע לן
Mai ka mashma lan · so what's the point?
What this is
A story in the Talmud doesn’t announce itself. It can be two sentences long or run across several pages. It often starts in the middle of a legal point. The same Hebrew words turn up in a ruling and in a story. Scholars learn to spot them over years of reading.
The goal is a complete catalogue of every story in every tractate, free for anyone to read in Hebrew and English side by side.
804
pages read so far, across four tractates
- Ketubot כתובות222
- Kiddushin קידושין162
- Gittin גיטין178
- Yevamot יבמות242
היכי דמי
Heikhi dami · what does that look like?
How it works
- א
Skim for events
Every sentence on every page is labelled: is something happening here, or is it argument? Pages with no events at all are set aside, which saves most of the work.
- ב
Ask which passages are stories
The AI reads each remaining page and marks the passages that are stories. It works from a written rulebook built from the expert's own rulings.
- ג
Look again, right next doorNew
The Talmud often tells the same kind of story two or three times in a row with different people. The AI tended to find one and stop. So a second pass looks at the passage beside every story it found and asks: is there another one here?
- ד
Set the edges
Each story gets a clear start and end, so it doesn't swallow the legal discussion around it.
- ה
Join what the page split
Page breaks in the Talmud are arbitrary. Stories that run across one are joined back together.
An example of step ג
Found first
The people of Upper Galilee buy a poor man of good family meat, every day.
Missed, right beside it. Now caught.
Hillel buys a poor man who had come down in the world a horse, and a servant to run before him. One day there is no servant, so Hillel runs ahead of him himself, for three mil.
פַּעַם אַחַת לֹא מָצָא עֶבֶד לָרוּץ לְפָנָיו, וְרָץ לְפָנָיו שְׁלֹשָׁה מִילִין.
Ketubot 67b. Both are lit on the page at the top.
מנא הני מילי
Mena hani milei · how do we know?
How well it works
As of October 2026
about
9in 10
of the stories on the expert's own lists are found
more than
9in 10
of the passages it calls a story really are one
about
4in 5
of its story start and end points land exactly where the expert put them
The second look is working
On Yevamot, adding step ג raised the share of stories found:
Before
89%
With the second look
~93%
80%100%
How we measured. We’d love to round these up. We didn’t.+
The expert, Prof. Jeffrey Rubenstein, wrote lists of the stories in these tractates in 2005, twenty years before this tool existed. So the first number can’t be flattered by anything we did: it is the share of his 452 listed stories that the tool finds, across four tractates (from 83% on Kiddushin to 97% on Gittin). The second comes from his reviews of passages the tool proposed. The third compares start and end points with those same 2005 lists. These are not yet as good as we want, and the tool isn’t perfectly repeatable: two identical runs can differ by a story or two in a hundred. So we compare runs in pairs before believing a change.
הלכה כמותו
Halakhah kemoto · the ruling follows him
How an expert keeps it honest
Prof. Jeffrey Rubenstein of New York University, one of the world’s leading scholars of Talmudic stories, reviews what the tool finds. He reads each passage in Hebrew and English and answers a few short questions. Is this a story? Does it start and end in the right place? Is it one story or two?
His rulings become a written rulebook, in his own words. When he decides a hard case, the decision is written down and the AI works from it from then on.
From the rulebook
A bare report of something isn’t a story. But:
“A man stole another man’s cow and sold it. Rava ruled… In this case you may have a story.”
Next: two AIs, one expert
Two different AI models
judge every passage against his rulebook
Where they disagree
plus a random sample, to keep them honest
The expert decides
a few passages, not the whole Talmud
The first test was a no-go. The models disagreed with him in a pattern that pointed to a rule missing from our rulebook. It’s added now, and the test runs again next.
“The AI was confusing attribution with characters. When it sees ‘Rabbi X said that Rabbi Y said…’, it thought there was a story with characters, but it’s just legal attribution.”
תיובתא
Teyuvta · the Talmud's word for “refuted.” We've earned a few.
What we got wrong along the way
We were measuring the wrong thing.
96% agreement
For months the headline number was “the expert agreed 96% of the time.” That only tells you how good the tool's picks are, not how many stories it missed. The expert's 2005 lists, written long before the tool, let us measure the misses. It's a humbler number, and the honest one.
Our test was too generous.
The way we matched the tool's stories to the expert's gave credit for a nearby passage, not the right one. When we tightened it, two scores fell by about 7 points. The tool hadn't changed; the scores were just finally accurate.
We took back our own best result.
We once credited the skimming step with the biggest improvement in the whole project. Re-checking it, we found the experiment had been set up wrong. We withdrew the claim.
The same run twice doesn't give the same answer.
Two identical runs on Yevamot disagreed on 3 of 102 stories. A one-story change between two runs means nothing, so now we always compare in pairs.
חברותא
Ḥavruta · study partners
The team
Scholarship & validation
Jeffrey L. Rubenstein, Ph.D.
Skirball Professor of Talmud and Rabbinic Literature, New York University
- Talmudic Stories: Narrative Art, Composition, and Culture (1999)
- Rabbinic Stories (2002)
- The Culture of the Babylonian Talmud (2003)
- Stories of the Babylonian Talmud (2010)
Technical development
Simon B.
Not a software engineer, but plays one on TV, with Claude Code as co-star.
Built with
- Sefaria: the free library of Jewish texts, Hebrew and English aligned sentence by sentence
- Google Gemini: the AI that reads the pages
- Anthropic Claude: the second AI judge, and the development assistant