How this study works

This page explains, from the ground up, how the Esther study on this site was done, what its results mean, and what they do not mean. It is written for a scholar who has never worked with software of this kind. Nothing here assumes you know what a “pipeline” or a “language model” is; both are explained when they first appear. Two sections near the end answer the questions a Talmudist and a Classicist are likely to ask, since they will not ask the same ones.

The whole study on one page

The question. The Greek Bible’s Esther is longer than the Hebrew. It carries six passages the Hebrew (the Masoretic text) does not: Mordecai’s dream, two royal letters, the prayers of Mordecai and Esther, Esther’s audience with the king, and the dream’s interpretation. Scholars call them the Additions to Esther, lettered A to F. The study asks whether any of this material survives in rabbinic and medieval Jewish literature: whether a Jewish text quotes it, retells its details, or shows no knowledge of it at all.

The material. The Additions were cut into 80 short units, each a clause or a single detail. They were compared against 2,125 paragraphs from 26 rabbinic, targumic and medieval Jewish works, from the Babylonian Talmud and the Aramaic targums to Rashi and the late medieval Esther anthologies. We call this collection the bench.

The method, in brief. A computer search and three independent machine readings proposed every possible pair of a unit and a paragraph that might be related, erring on the side of too many. About 1,250 candidate pairs came out. Each was then read again, under strict written rules, and ruled verbatim (the wording is reproduced), motif (a specific detail is shared) or unrelated. Human checks and a hostile review measured how often those rulings were wrong, and a second round re-ruled every uncertain pair. Finally, each confirmed pair was traced back, where possible, to the version of Esther it came from.

The results.

• No verbatim quotation of the Additions in ancient rabbinic Hebrew was found on this bench. The passages that look like quotations, in Esther Rabbah 8:5, 8:7 and 9:1 and in Yalkut Shimoni, are medieval. They were taken from Sefer Yosippon, a tenth-century Hebrew chronicle that itself translated the Additions from the Latin Vulgate. They show the Additions coming back into Hebrew in the Middle Ages, not surviving from antiquity.
• One ancient text reproduces the wording of an Addition: Targum Sheni, an Aramaic targum of Esther, quotes the king’s second letter (Addition E) clause by clause at Esther 8:13. Its details match the Greek, or the Old Latin translated from the Greek. They do not match the Vulgate, Josephus or Yosippon.
• 136 pairs share a motif, a specific detail the Hebrew Esther does not have. Most gather where scholars have long looked: the Talmud’s account of Esther’s audience with the king (b. Megillah 15b–16a), Esther’s refusal of the king’s food, and the reasons Mordecai would not bow. Seventeen of these pairs are in works older than Yosippon.
• Where the motifs came from cannot be told from the motifs. The details they share are, almost always, in several versions of Esther at once, so a shared detail rarely points to one version.
• 34 of the 80 units have no rabbinic witness outside the medieval borrowings: chiefly the royal letters’ official framing, the dream and its interpretation, and the plot of the two eunuchs.

That is, in the main, a negative result. It is published as one.

Part one: the pieces, explained

1. Why the Greek Esther is the starting point

When a book of the Greek Bible contains material the Hebrew lacks, there are two broad explanations. Either the Greek translator, or someone after him, added it; or the translator had a Hebrew (or Aramaic) text that contained it, and that text was later lost. Since the Dead Sea Scrolls showed that several forms of biblical books circulated in Hebrew in antiquity, the second explanation has had to be taken seriously for every such passage.

If a lost Hebrew or Aramaic form of Esther once contained the Additions, it might have left traces in Jewish literature, which kept reading and retelling Esther. So the Greek serves here as a finding aid: it tells us what to look for. The rabbinic corpus is the place where we look. The study does not assume the answer. It asks where the traces are, and how each one got there.

2. The units

The Additions were not searched as whole passages. A rabbinic text rarely retells a whole prayer; it retells a detail from it. So each Addition was cut into units small enough that a rabbi could have retold one of them on its own: one clause, one image, one claim. That produced 80 units.

The Greek text used throughout is Swete’s edition of the Septuagint (the “B-text”). To find where the Greek and Hebrew differ, the project used the CATSS parallel alignment of the Greek and Hebrew Bible, as an index only. Each unit carries a draft English translation made for this project, marked as a draft wherever it appears.

3. The bench

The bench is the body of Jewish texts searched: 2,125 paragraphs from 26 works. Each paragraph carries its source, its licence and a date. The main works are Targum Sheni and Targum Rishon on Esther, the Babylonian Talmud’s tractate Megillah, Esther Rabbah, Pirkei de-Rabbi Eliezer, Rashi on Esther and on Megillah, Panim Aherim, Abba Gurion, Yalkut Shimoni and several medieval Esther anthologies. A few works are there as controls (explained below): texts that should not show the Additions, to test whether the search invents parallels.

Dates are given per stratum, not per work. Some works were composed in layers centuries apart. Esther Rabbah is the important case. Its first part (the proems and chapters 1–6) comes from the land of Israel, no later than the early sixth century; its second part (chapters 7–10) was edited around the eleventh. The study first dated the whole work as one, and that mistake nearly produced a false headline, because the strongest parallels are all in the late part. Each paragraph is now dated by the layer it belongs to.

4. What a “pipeline” is

A pipeline is simply a fixed sequence of steps, each of which takes the previous step’s output as its input, and each of which is written down so that it can be run again and give the same result. Here the steps were: cut the units; assemble the bench; search; read; verify; rule; check. The point of working this way is that every number on this site can be traced back through the steps to the texts, and every step can be inspected, criticised and repeated.

5. What an “agent” is

The machine readings in this study were done by language models, the kind of software behind tools such as ChatGPT or Claude. A language model reads text and writes text. It can read Hebrew, Aramaic, Greek and Latin, with a competence that is real but not infallible.

An agent is one such model given a specific job: a written brief, a set of files to read, and a fixed form in which to return its answers. For example: “Here are the 80 units and 40 paragraphs of Targum Sheni. For each pair, rule verbatim, motif or unrelated, by these definitions, and give one sentence of reason.” Several agents can work at once on different parts of the material. Each works only from its brief and its files; it has no memory of other runs.

Two practical consequences follow. First, the brief matters enormously: an agent does what its brief says, including what it says badly. The briefs were therefore written down, revised when they failed, and kept with the results. Second, an agent’s reply can disagree with its own output. So the study never counts from what an agent said it found; it counts from the files the agent wrote.

6. Recall first, then verify

There are two ways a search can fail. It can miss a real parallel (a failure of recall), or it can report a false one (a failure of precision). No single step does both jobs well, so the study used two stages.

Stage one casts a wide net. Its only job is not to miss anything.

• A lexical search. For each unit, two to four possible Hebrew wordings were composed (a retroversion: a guess at what a Hebrew text behind the Greek might have said). A program then looked for paragraphs on the bench that share runs of letters with those wordings, giving extra weight to rare words and little to common ones.
• Three machine readings. Three sets of agents read every work against every unit, each under a different brief. One started from the units’ details and looked for them in the texts; one took each unit in turn and asked whether the text had anything like it; one was told to assume that the parallels the scholarship already knows were wrong, and to look for what else might be the real echo. They were told to keep anything plausible.

Together these produced about 1,250 candidate pairs, where the study had expected perhaps 40 to 120. That is what a wide net is supposed to do.

Stage two sorts the catch. Each candidate pair was read again, against its unit, and ruled under written rules (section 7). A candidate was to be rejected if the shared material is already in the Hebrew Esther; if it is a stock biblical phrase; if a long paragraph matched on vocabulary alone; if a shared word refers to something different; if the parallel is better explained by another biblical book; if the passage contradicts the unit; or if it is merely a copy of another witness.

7. Verbatim, motif, unrelated

These three words are used in strict senses.

• Verbatim: the witness reproduces two or more of the unit’s clauses, in the unit’s order, with the same content, and renders at least one specific word of the unit in a way that is neither the Hebrew Esther’s wording nor a stock phrase. The language does not matter: an Aramaic rendering can be verbatim. A paraphrase that keeps the order and content across three or more clauses qualifies.
• Motif: the witness shares a specific detail with the unit that the Hebrew Esther does not supply. “Specific” means a name, a number, a place, an object, an image, an action, a speech or a distinctive epithet. It does not mean a theme (“divine deliverance”), a literary device or a narrative function.
• Unrelated: neither.

8. Why precision had to be measured, not assumed

A ruling made by a machine, or by a person, is a claim. Before a count of rulings means anything, we need to know how often such claims are wrong. That can only be found out by checking a sample by hand against the same rules.

This mattered here. The editor first read forty paragraphs whole and overturned none of the machine’s rulings. A separate adversarial review (a second, independent model instructed to attack every claim the study made) then took a random sample of forty individual pairs and applied the written rules strictly. It rejected 58% of the motif pairs. Both checks were honest; they measured different things. The editor asked whether each paragraph belonged in the study. The review asked whether each pair met the rule. A paragraph can belong while most of its pairs fail.

So the motif tier was re-ruled in a second round. First the rules were sharpened with worked examples of the three commonest errors. Then the rules were tested: a fresh reading of the same forty pairs had to agree with the review’s rulings on at least 34 of them before anything else was run. It agreed on 35. Only then were all 509 uncertain pairs re-read, each blind to its earlier ruling. 366 (72%) were rejected. That left 72 verbatim and 136 motif pairs, and 1,041 unrelated.

9. Controls

A control is a test whose answer is known in advance. Before any search was run, the study fixed a list of pairs from the scholarly literature:

• Positive controls are parallels the literature already knows (for example, that Esther Rabbah 8:5 retells Mordecai’s dream). A good search must find them. It found all of them that were on the bench, including Targum Sheni 8:13, which the model had failed to name when asked, from memory alone, where the Additions appear in rabbinic literature.
• Negative controls are places the literature says the Additions are absent (for example, that Targum Sheni does not retell Mordecai’s dream). A good search must not “find” anything there. These are the only real test of precision at the level of motif, because motifs are easy to imagine.
• Machinery controls are texts that are certainly translations of the Additions (a medieval Hebrew rendering of them, for instance). If the search cannot recognise those, it is broken.

10. Conduits and date gates

Finding a parallel is not the end. The next question is how the material got there. The study calls the route a conduit, and uses a small fixed vocabulary:

• medieval-retranslation: the material came back into Hebrew in the Middle Ages from a Latin or Greek text (Yosippon is the main example);
• aramaic: it is carried by an Aramaic targum;
• oral: shared narrative tradition, with no sign of a written source;
• lost-intermediary: a written source that no longer survives;
• greek: contact with a Greek text;
• hebrew-vorlage: a Hebrew text behind the Greek (the word Vorlage means the text a translator worked from).

Dates rule some routes out. A text written before about 400 cannot depend on Jerome’s Vulgate, which did not yet exist. A text written before about 953 cannot depend on Sefer Yosippon. These are the date gates. They are why per-stratum dating matters: a paragraph in Esther Rabbah’s early part cannot come from Yosippon; a paragraph in its late part can.

11. Translation fingerprints: what the second round did

The last step asked, for each confirmed pair, which version of Esther the shared detail came from. There are several: the Septuagint (the B-text); a second Greek form called the Alpha-text; the Old Latin, translated from Greek before Jerome; Jerome’s Vulgate; Josephus’s retelling in his Jewish Antiquities; and Sefer Yosippon’s Hebrew. They do not all tell the story the same way. A detail that only one version has is that version’s fingerprint. If a rabbinic text shares it, that points to the version.

To make this possible, every one of the 80 units was broken into its details, 375 of them, and for each detail each version was checked: present (with a quotation), absent (after searching the whole text), or present in a different form. Another 188 details were recorded that one version has and the Septuagint lacks. Every quotation, 1,916 of them, was checked by computer against the source text. The Alpha-text, the Old Latin and Yosippon were transcribed for the purpose from printed editions (Lagarde 1883, Sabatier 1743, Breithaupt 1710).

Then five agents read the 136 motif pairs and Targum Sheni’s letter, mixed blind with 50 controls: 20 pairs whose route is known (ten from Yosippon, ten from a medieval Hebrew translation of the Vulgate) and 30 pairs already ruled unrelated. The agents saw the witness texts under code numbers, not under their titles, so they could not use what they knew about a work. The method passed its test: all ten Yosippon controls pointed to Yosippon; nine of the ten Vulgate controls pointed to the Vulgate, and the tenth pointed nowhere; none of the thirty negatives was assigned a route.

What it found. Of the 136 motif pairs, 80 share a detail that the Hebrew Esther lacks. Ten point to a single version, all of them Yosippon, and the strong ones are in the paragraphs already known to borrow from Yosippon. The other seventy point to several versions at once, most often the two Greek texts and the Vulgate together. None points to the Greek alone or the Latin alone. In 48 of the 136 confirmed pairs the fingerprint readers found no specific shared detail at all: a second, independent measure of how strict the motif tier still needs to be.

What it proves, and what it does not. A fingerprint can show that a rabbinic text knew a detail that is in the Additions and not in the Hebrew Esther. It can show that a medieval text took its wording from Yosippon or from the Vulgate. It cannot, for most motifs, show which version a rabbi knew, because the versions share most of the Additions’ details. It says nothing about how a detail travelled: by reading, by hearing, or through a lost intermediary.

12. The answer to the study’s question

Did rabbinic literature know more of the Greek and Latin traditions of Esther than is usually credited? The count is this:

• Wording: one ancient witness, Targum Sheni 8:13, which follows a text of the Greek type (in Greek, or in the Old Latin made from it) and not the Vulgate, Josephus or Yosippon. This was checked against the critical edition of the Old Latin (Haelewyck, Vetus Latina 7/3, 2003–2008): the decisive detail, “innocent blood” in the king’s letter, is in every Old Latin text-type that has the verse and in no manuscript of the Vulgate.
• Motif: 136 confirmed pairs; 80 share a detail the Hebrew lacks; 17 of those are in works older than Yosippon (the Babylonian Talmud, Targum Rishon, Pirkei de-Rabbi Eliezer and the early part of Esther Rabbah). They show that the content of the Additions, above all the prayers and the audience scene, was known in rabbinic tradition, including before the medieval route existed. They cannot show which Esther carried it.

The motifs are notable as reception of the Additions’ content. Their provenance will not be settled from this direction. The Targum Sheni letter is the study’s clearest finding.

13. Limits

• The bench is not all of rabbinic literature. Lekach Tov on Esther and Aggadat Esther are not on it, nor is Sefer Yosippon itself as a bench text. A unit with no witness has none on this bench.
• The texts are printed editions, not manuscripts, taken from Sefaria and Hebrew Wikisource at fixed versions.
• The comparison editions are older than the best ones: Lagarde’s Alpha-text rather than Hanhart’s Göttingen edition; Sabatier’s one Old Latin manuscript rather than Haelewyck’s edition (except for the Targum Sheni readings, which were checked); Breithaupt’s Yosippon rather than Flusser’s critical text, which was consulted in Bowman’s English translation.
• Machine reading has an error rate, which the study measured rather than assumed, and which is reported with each result.

Part two: for the Talmudist

Which editions were used, and why? The texts come from Sefaria and Hebrew Wikisource, at pinned versions, because they are openly licensed and can be republished beside the results. Each paragraph on the site names its edition. In brief:

work	edition used	date assigned
Babylonian Talmud, Megillah	Vilna text (via Wikisource)	c. 200–500, edited later
Targum Sheni on Esther	Berlin 1898	c. 600–1000, composite
Targum Rishon on Esther	Mikraot Gedolot	c. 400–800
Esther Rabbah	Hebrew Wikisource, Vilna pagination; chapter 9 from Tosefet le-Esther’s copy and the Torat Emet text, collated against Vilna 1884	part I c. 400–600; part II c. 1000–1150
Pirkei de-Rabbi Eliezer	Sefaria’s vocalised edition	c. 750–850
Panim Aherim, Abba Gurion	Buber, Vilna 1886	c. 850–1300
Yalkut Shimoni	Sefaria’s text	13th century
Rashi on Megillah; on Esther	Vilna; Metsudah	1040–1105
Midrash Megillah and the other Esther pieces	Eisenstein, Otzar Midrashim (New York 1915)	as each piece’s source allows

These are not critical editions, and the study does not pretend they are. The Tabory–Atzmon critical edition of Esther Rabbah (Schechter Institute, 2014) has not yet been consulted.

How was Esther Rabbah’s dating established? Not by assumption. The division into an early part (proems and chapters 1–6, from the land of Israel, no later than the early sixth century) and a late part (7–10, about the eleventh century), and the late part’s borrowing from Yosippon, are stated by Theodor (Jewish Encyclopedia), Herr (Encyclopaedia Judaica) and Atzmon (Jewish Studies, an Internet Journal 6, 2007). The borrowing was then checked against Yosippon’s own text: in Breithaupt’s printing (Gotha 1707, reissued 1710), twelve phrases of Esther Rabbah 8:5 recur in the same order, including a frame that the Greek does not have; and details that Yosippon took from the story of Joseph and Aseneth (Esther “begging from house to house, from window to window”; God “father of orphans”) recur in Esther Rabbah 8:7 and 9:1. Those details are in no version of Esther, so they can only have come through Yosippon.

Can a machine be trusted to read Rashi or Targum Sheni? Not on its word, and the study never takes its word. What was checked is the ruling, not each word of the reading. The machine’s rulings were tested against known answers (the controls), against a hostile review, against a hand-scored sample that a fresh reading had to reproduce before any volume work began, and, in the second round, by re-reading every uncertain pair blind. Where the study transcribed printed pages itself (Yosippon, the Old Latin, the Alpha-text), samples were re-read against the page images; the error rate on the poorest Hebrew page was about one word in 260. Every paragraph is published beside its ruling, so any reader can check any pair.

What do “verbatim” and “motif” mean in terms you would recognise? Verbatim is the relation of two texts that share a passage, the way a baraita appears in two sugyot, or a targum renders its verse: the same clauses, in the same order, in recognisably the same words, whatever the language. Motif is the relation of two texts that know the same aggadic detail without sharing wording: the Talmud’s three angels who attend Esther before the king, set beside the Greek’s God who “changed the king’s spirit to gentleness”, is the kind of thing the study means. A motif shows knowledge of a story; it does not show knowledge of a text.

Why does the Talmud’s material count as “oral”? Because its parallels are shared details, not shared wording, and because the Babylonian Talmud is too early and too indirect to argue that its rabbis read a written form of the Additions. “Oral” here means shared narrative tradition. It is a ruling of the study, stated with its confidence, and open to challenge.

Part three: for the Classicist and the Near Eastern scholar

Which Greek? Swete’s Septuagint (the B-text) for the units, because it is openly licensed and can be published; the Göttingen edition (Hanhart) is in copyright. The Alpha-text is Lagarde’s (1883), transcribed from the page images for this study. The CATSS alignment was used as an index only.

Which Latin? The Vulgate’s Additions are in the appendix where Jerome placed them (10:4–16:24), from the Clementine text. The Old Latin is Sabatier’s (1743), whose base for Esther is the Corbie codex, a single manuscript. Haelewyck’s critical edition (Vetus Latina 7/3) sorts the Old Latin manuscripts into several text-types; it was used to check the readings on which the Targum Sheni result rests, and it corrects two apparent differences between the Greek and the Old Latin that were only slips in Sabatier’s manuscript.

How does the relationship between the Alpha-text and the B-text bear on what a fingerprint can prove? Directly. The two Greek texts differ considerably in the parts of Esther they share with the Hebrew, and how they are related is disputed: one view makes the Alpha-text a translation of a different Hebrew Esther, with the Additions attached from the Septuagint; another makes it a Greek rewriting of the Septuagint. Jobes (The Alpha-Text of Esther, 1996) and De Troyer (Rewriting the Sacred Text, 2003) are the standard treatments; the project has not yet read them. For the purposes of this study the dispute makes little difference, because in the Additions themselves the two Greek texts share nearly every detail. A rabbinic detail found in one is almost always in the other, so a fingerprint can rarely choose between them. That is why Targum Sheni’s letter is placed with “the Greek (B or AT)”, not with one of them.

Can the method distinguish a Greek source from a Semitic one behind it? No, and it is important to say why. The method compares a rabbinic text with the surviving versions of Esther. A lost Hebrew or Aramaic original of an Addition would, by definition, have contained the details its Greek translation contains. So a rabbinic text that agrees with the Greek is equally consistent with the Greek and with a Semitic text behind it. The only kind of evidence that could separate them is a rabbinic reading that agrees with a reconstructed Hebrew wording against the Greek, for instance where the Greek looks like a mistranslation of a Hebrew word and the rabbinic text has the right word. The study looked for such cases and found none: no pair carries the hebrew-vorlage route. For the royal letters, Additions B and E, a Semitic original is in any case widely doubted, since their style is that of Greek chancery prose.

What about Josephus and Yosippon? Josephus retells Esther in Greek, including four of the six Additions (not the dream or its interpretation) (Antiquities 11.184–296); the study used the Perseus text of Niese’s edition. His version differs from the Septuagint in places, which makes it a useful fingerprint: where Targum Sheni’s letter and the Septuagint agree against Josephus (twice: “innocent blood” and the Jews as God’s “sons”), Josephus is not the source. Yosippon is a tenth-century Hebrew reworking of Josephus that took the Additions from the Latin; in two recensions read (Breithaupt’s printing, and Flusser’s text in Bowman’s translation) it quotes neither royal letter, so Targum Sheni’s letter did not come from it either.

Reading still to do

These works bear directly on the study and have not yet been consulted. Nothing on this site should be read as engaging them:

• Tabory and Atzmon, critical edition of Midrash Esther Rabbah (Schechter Institute, 2014): the modern basis for dividing Esther Rabbah into layers.
• Saskia Dönitz on the manuscript recensions of Sefer Yosippon (in From Josephus to Yosippon and Beyond, Brill): at least four recensions exist, and Breithaupt’s and Flusser’s are not all of them.
• Karen Jobes, The Alpha-Text of Esther (1996), and Kristin De Troyer, Rewriting the Sacred Text (2003): the standard studies of the two Greek texts.
• Grossfeld’s and Ego’s editions of Targum Sheni, for the manuscript tradition of the 8:13 letter.

Where to check everything

Every unit has its own page, with its Greek, its draft English, and every witness paragraph ruled against it. The method page gives the procedure and the controls in brief; the findings page, the results; the negative results page, the units with no witness; the export page, the data itself. The records of the pipeline, the briefs and every check described here are kept with the project’s data.