Closing the Fragment: Sappho 31 and the Composition Layer as a Surface of Classical Reception (v8)

Closing the Fragment: Sappho 31 and the Composition Layer as a Surface of Classical Reception (v8)

Sappho 31 and the Composition Layer as a Surface of Classical Reception

Lee Sharks · v8 · 7 November 2026 · Machine-Mediated Reception Studies · rebuilt in place from v7 (#1749), which was revised from v6 (CLASSREC-2026-062, declined 7 November 2026)

Abstract

On 18 January 2026 Google's AI Overview, asked about Rebekah Cranes's translation of Sappho 31, returned a fifteen-line poem headed "Rebekah Crane's Translation", ending "But I must dare all.", and told the reader that the κῆνος of the first line is "you, the future reader, who completes the poetic circuit." The one source the answer displayed was the translator's book on Medium, whose text of the poem ends "But everything must be endured, since even a poor man…" under a heading that sets κῆνος equal to the future reader. Within the interpretive environment those documents built, the composition produced a conclusion none of them contains, the poem closed in the first person and set under its source's gloss, and published it under the translator's name with no mark of what it had added.

The article takes that act as an act of reception. An answer engine selects a text, presents it, assigns it a name, explains it and represents the standing of interpretations of it; each is an editorial and critical act, and the record observes each one. The question is what this act does to this poem. On the conventional reading the poem breaks at τόλματον, and the composition closed a broken text. On the reading the archive has argued since August, the transmitted poem ends between two open positions, the unoccupied seat at κῆνος and the empty agent of τόλματον, and the openness is its form. The composition repeated its source's reading of the seat, the reader as κῆνος, and joined to it a determination of the agent, "I must dare all", that no document in its environment makes. An occupation of an open position through reading, and a determination of the text attributed to its translator, arrive in one answer under one name. The transmission history, read from seated texts ('Longinus', Catullus 51, Bergk 1882, Edmonds 1922, Wharton, and English versions from John Hall in 1652 to 2013), shows each receiver deciding what kind of completeness the poem may have, and the translators mostly closing it without a mark. The output's wording corresponds to that history, to Wharton's literal version above all; correspondence locates antecedents and shows no route of composition. Closure is defined as three operations and signature at three levels, the page, the edition and the network. The finding: the composition layer receives a poem inside the environment its documents make, decides the poem's completeness, and transmits the decision under another's name. A dated series of compositions to October 2026 and a protocol for frequency, variation and causal sensitivity extend the analysis.

Keywords: Sappho; Fragment 31; classical reception; translation history; κῆνος; τόλματον; open position; Wharton; Bergk; Edmonds; Catullus; apparatus criticus; obelus; AI Overviews; machine-mediated reception; capture

0. Note on this version

v6 was declined by Classical Receptions Journal on 7 November 2026 on two referee reports. v7 (#1749, the same day) answered every point in them (Appendix A) and a further review (Appendix B.1–B.4). Its revisions were of two kinds. Corrections of record stand in this version: the June frames and figure numbers, the dates of the translator's texts, the English baseline, the first printings seated as EA-CORPORA-36 (#1748). A second kind of change went past the evidence. v7 read the correspondences between the output's lines and Wharton's literal version as the route of composition, a compilation of two texts, and withdrew v6's finding that the output was composed. The record of the answer's displayed source (#337 §5.2) shows no Wharton, and a correspondence of wording establishes no source. v8 holds the alignment as correspondence (§5.3), restores v6's finding in corrected form, sets the January answer against the environment its displayed source built, and puts the reading of the edge as the poem's form at the centre of the argument (§6.6, §7). Appendix B.5 records what changed.

1. A fragment that came back whole

Sappho Fragment 31 survives because 'Longinus' quoted it (On the Sublime 10.2) as an instance of the sublime. Four stanzas come through. The fourth ends in the body's failure: τεθνάκην δ' ὀλίγω 'πιδεύης / φαίνομ', "I seem to myself little short of dying". After it the Longinian text carries ἀλλὰ πᾶν τόλματον, "but all is to be dared", and then words whose attribution to Sappho is disputed, ἐπεὶ καὶ πένητα. Editors differ in how much of this they print as Sappho's and how they mark the edge (Voigt 1971; Campbell 1982).

The conventional reading treats the edge as a loss, and the edge has been handled in two broad ways. Editors and literal translators have marked it: Bergk with brackets, Wharton with an ellipsis, and Carson in 2002 with an ellipsis that her book made famous. Translators who wanted an English poem have most often closed it: Behn (1694) ends "I sigh, I tremble, and I dye"; Ambrose Philips (1711) "I fainted, sunk, and dy'd away"; Merivale (1813) "Both sight and sound"; Symonds (1883) adds "Lost in the love trance"; William Carlos Williams (1957) ends at the fourth stanza; Lowell (1961) ends "I die!"; Barnstone (1965) closes the transmitted continuation into a sentence, "but / I must suffer everything, being poor." Carson's own version of 1986 ends at the fourth stanza. Of thirteen English versions from 1652 to 2013 in one teaching gallery, eleven present the poem closed and without a mark as the gallery reproduces them, and twelve when Hall is read in his own print (§6.4). The gallery began as a double-sided sheet of translations through Carson prepared by Yopie Prins, whose Victorian Sappho (1999) is the study of that English reception; the author extended it and taught from it in secondary classrooms. Until Carson's ellipsis of 2002, a reader of Sappho in English most often met the poem closed.

On 18 January 2026, asked about the Cranes translation, Google's AI Overview returned:

That man seems to me a god on a par with the gods,

…

I am greener than grass.

I feel I am all but dead.

But I must dare all.

It headed the poem "Rebekah Crane's Translation (Fragment 31, "That Man")", displayed one source, the translator's book on Medium, and named it: "as presented in the Crimson Hexagon document "Day and Night: Conversations with Sapphic Desire"". It then stated, unprompted, that the reader of the translation was the addressee of the poem's first word, the κῆνος, "that man", and that reading completed a transmission twenty-six centuries old.¹

¹ Rebekah Cranes is a heteronym under which the present author publishes. The disclosure comes at the first mention because it is a condition of the method: the comparison in §5 requires the translator's texts in their dated, hashed states with their version history, and only the archive that holds them can supply that. Its consequences for the design are set out in §9.

The answer closed the poem on a line its source does not carry, and stated a reading of the poem's first word in the same breath. The archive has argued since August that the transmitted poem is complete with two positions open, the seat at κῆνος and the agent of τόλματον (§6.6). The January answer repeated its source's reading of the first and determined the second.

This article reads that event as reception and as a recorded observation, and holds each reading to its own standard of evidence. Reception studies has long held that each receiving culture constitutes what it receives. An answer engine that selects a text, presents it whole, names its translator, explains it and states its standing has received it, in public, for a reader who takes the answer as an answer about antiquity. Those acts are on the record. An experiment can investigate why a given presentation occurs; the record already establishes that it occurred. The field has instruments for receptions that leave artifacts: translations, editions, adaptations, performances. A composed answer is produced for one reader and is gone. A capture makes it a dated object, and a registry of captures makes a series of them comparable.

2. Terms

Composition layer. The stage of an answer engine that writes prose from retrieved material, distinguished from the index that finds and ranks pages. Google describes AI Overviews as generated answers grounded in Search results; the composition layer is that writing stage, observed from outside.

Closure. The presentation of a text as ending securely at an edge its transmission leaves disputed. Whether the edge of fragment 31 is a break or an achieved openness is a question of reading (§6.6); the operations are observable either way. Three operations produce closure, alone or together:

• added continuation: words are supplied beyond the transmitted text (Smollett's stanza; Symonds's last phrase; the Cranes reconstruction);
• terminal selection: the text is stopped at a chosen point and presented as ending there (Philips and Merivale at the fourth stanza; Barnard; Bergk's withdrawn conjecture at verse 16);
• resolution of suspended syntax: a construction the Greek leaves open is closed in the rendering (τόλματον, "to be dared", given a subject: Wharton's "I must dare all").

Signature. Any mark that tells a reader which words are transmitted and which are supplied, or that the text stops short: an ellipsis, brackets, a note, a heading, a declared omission, a change of voice. A signature is assessed at one of three levels, and the level is always stated:

• the page: what a reader of that presentation can see;
• the edition: the presentation with its own apparatus, preface and notes;
• the network: the documents a determined reader could reach from the presentation, including its source links.

Retrieval basin (the author's coinage). The set of indexed surfaces from which the composition layer draws when asked about a given entity. It cannot be inspected. This article estimates it from what had been published by a given date, from source chips and cards shown with an answer, and from the organic results on the same page, and says which of these each claim rests on.

Capture. A documented composition: query string, surface, sign-in state, date, the output transcribed verbatim, a frame where one was taken, and an annotation. Address. A fixed query string re-run over time.

Interpretive environment. The documents a composition is observed to draw on, with the designations, relations and readings they establish. It is estimated from the sources an answer displays and from what the answer reproduces of them, and it is narrower than the basin.

Open position. A place in a text's form left without its filler. Fragment 31 has two: the seat opposite the addressee that κῆνος holds, and the agent that the verbal adjective τόλματον leaves unexpressed.

Position as attractor (the author's coinage; §7). The proposal that an interpretive position established in the environment governs which configuration is composed, and under which name.

3. From classical reception to machine-mediated reception

Jauss's horizon of expectations (1982: 22–3) names the norms and prior readings that shape what a reader can find in a work; Iser's implied reader (1978: 34–8) names the textual structures that anticipate the finding. Budelmann and Haubold (2008: 13–15) set reception against tradition: where tradition stresses continuity, reception foregrounds what the receiver does. Martindale (1993; 2006) carried this into classics, and Hardwick (2003) gave the field its primer.

The receiver studied here performs, in public, acts the discipline already names: it selects a text, presents it as a whole, attributes it, explains it, and represents the standing of an interpretation ("famously read as"). An editor does each. The study takes these acts as given in the record and asks what they do to the poem. The receiver is a distributed apparatus: indexed pages, ranking, the query, computational synthesis, interface conventions, and the reader who takes the answer as an answer about antiquity. The capture observes the event that apparatus produces; nothing here claims knowledge of a model's interior. In a limited and exact sense the apparatus has a horizon: what is indexed and what ranks determines what can be found. It has conventions of presentation: composed prose, a confident register, bullets, source chips. Its observable relation to antiquity runs through existing reception surfaces: in this case the translator's book its answer displayed, and an English idiom of the poem that runs back to Hall. Its receptions do not persist. Platform scholarship (Gillespie 2014; Noble 2018) established that ranking shapes what is knowable; this article asks what such an apparatus does with a text whose transmitted form has no ending.

4. Design

The study has four parts. It reads one documented reception event closely, against the environment its displayed source built (§5.1–5.3). It reads what the event did to the poem, under the conventional reading of the edge and under the reading of the edge as form, set in the poem's editorial and translational history, read from seated texts (§6). It follows the poem and its translator through a dated series of compositions from June to October 2026 (§5.4–5.5). And it states a protocol that extends the analysis (§8). The claims stand at four levels.

Documented. On 18 January 2026 the AI Overview, displaying one source, the translator's book on Medium, returned under the translator's name a fifteen-line poem ending "But I must dare all." and a gloss naming the reader as the κῆνος (§5.1). The source's text of the poem ends on the agentless "But everything must be endured" and an ellipsis, under a heading that sets κῆνος equal to the future reader (§5.2). The output repeats the source's assignment of the seat to the reader, supplies the agent the source leaves empty, and carries no mark (§5.3). The record is a same-day verbatim transcription; no frame was taken.

Read. What the act does to this poem: under the conventional reading, a broken text closed and its mark dropped; under the reading of the edge as form, a completeness decided for the poem and published as the translator's (§6.6).

Proposed. Why this configuration and this name: position as attractor (§7).

Extended. How often the composition layer closes the open positions of fragmentary texts, how the closure varies, and what in the environment it responds to (§8).

5. The record

5.1 The January event

The 18 January output survives as a verbatim transcription deposited the same day in Alexanarch record #337, "The Summarizer Becomes Translator", with the query context and the translator's text the output names. No frame was taken, no third party witnessed it, and it cannot be re-run. It is reported at that standard.

Two features of the transcription bear on the June series. The output is headed "Rebekah Crane's Translation", so the collapse of the heteronym into the name of an existing novelist, which v6 first dated to 14 June, is present on 18 January. And the output names its source document correctly. The answer displays one source, "Day and Night: Conversations with Sapphic Desire" on Medium ("1 site"). The session's earlier answers display five more: two versions of a navigation map, an audit protocol, the book Pearl and Other Poems on Amazon, and "Various DOI-anchored documents" on Zenodo (#337 §5.2). Wharton is among none of them.

The event precedes a dated platform change. On 27 January 2026 Google announced that it was "making Gemini 3 the new default model for AI Overviews globally" (Google 2026). The January capture was taken before that announcement and the later Overview captures after it. Google does not report which model wrote a given answer, so the announcement dates a change of default and says nothing about any single answer. The comparisons in §5.4–5.5 run across that change and are read as longitudinal.

5.2 What had been published by 18 January

v6 cited the translator's texts by records dated 2 February (#436) and 16 March (#1167), both later than the event. The comparison is now made against the deposits published by 18 January. Publication is the finding; whether a given deposit had been indexed by that date the record cannot show. The output's naming of Day and Night is direct evidence of its relation to that document.

The environment. The answer's displayed source is the translator's book on Medium, and #337 quotes its text of the poem the same day under the heading the book gives it, which ends: 'κῆνος = "that man there" = the future reader = you.' The session's answers describe the documents they draw on as designating that book the "Ground Truth" for the poem, and state their claim: 'The "circuit completes" in your processing of the text'. Every rendering of the dare-line the archive had published by 18 January leaves τόλματον without an agent: the book's "But everything must be endured", the reconstruction's "But all must be dared" (#281) and "All may be dared" (#1270, #313).

The translation. Day and Night was deposited on 9 January with the DOI 10.5281/zenodo.18202658. Its record lists two files, DAY_AND_NIGHT.md and DAY_AND_NIGHT.pdf, and its description says of the translations "Gaps and ellipses retained". Record #337 quotes the text the same day: its Sappho 31 opens "He is an equal to the gods, that man" and ends "But everything must be endured, since even a poor man…". The files were lost with the repository account (§9), and the record cannot confirm whether both carried that rendering. The operator recalls the deposited book carrying the reconstructed version of the poem. The Day and Night blog post, as recovered on 15 October 2026 (#324), carries both: the ellipsis rendering in its text, and on its book pages a rendering that opens "That man seems blessed as any god to me" and ends with the reconstructed stanza, "let this song survive in script", with no notice on that page. The 2013 book text (record #282) ends the poem at the fourth stanza, "I think I'm close / to the final darkness."

The reconstruction. The lost stanza was reconstructed and published with its warrant on 9 January: in "ΦΑΙΝΕΤΑΙ ΜΟΙ: Sappho 31 and the Inscription of the Future Reader" (#1270, §V.3), whose Zenodo record (#281) gives the Greek and an English version under the heading "Reconstruction of the Lost Fourth Stanza":

ἀλλὰ πᾶν τόλματον· ἐπεί σε, Κύπρι,

καὶ πένητά γε κἄ βασιλῆα δάμασσε·

καὶ πόλεις ὀλβίους ἀπέθηκας ἄφαντον·

γράμμασι μολπὰν

"But all must be dared; for you, Kypris, have yoked beggar and king alike, and you have cast down even prosperous cities to nothing—let the song survive in letters."

On 14 January "The Kenotic Truth of Sappho 31" (#313) republished it as "The Reconstructed Final Stanza" and stated its making in its own apparatus: "MACHINE_MEDIATED_SCHOLARSHIP", with participants "Human_Philologist" and "Assembly_Evaluation". (The stanza was then called the fourth; the erratum of 14 June, #201, renumbered it the fifth, since four stanzas survive.)

The interpretive claim. Published by 9 January: that κῆνος addresses a future reader who completes the poem's transmission (#1270; #283; #313).

The literal translation. Wharton's Sappho, with its literal prose version of fragment 31 (his fragment 2), is in the public domain and transcribed on Project Gutenberg (#57390, fifth edition, 1908; first edition 1885). It is a comparison text for §5.3; the record shows it among none of the answer's sources.

5.3 What the output is, and what it did

The output can be set against the text its displayed source carried and against the English tradition. Aligned line by line (Appendix C), five of its fifteen lines correspond most closely to the translator's January text, six to Wharton's literal version, one to Carman's version of 1904, and three to none of these in a run of three words or more. A correspondence of wording records no source and no step of composition: shared wording arises wherever renderings of the same Greek are made under the same constraints of grammar, idiom and inherited English, and Wharton's version is itself the product of such constraints and of his inferences, as every translator's is (§6.4). The correspondences place the output's language. Its body runs with the text its source carried; its first and last lines run with the oldest English idiom of the poem, Wharton's "peer of gods" modernized and the dare-line from Hall through Wharton and O'Hara.

What the output did can be stated against its environment, and the statement needs no route. Its source ends the poem "But everything must be endured, since even a poor man…": τόλματον rendered without an agent, the disputed words ἐπεὶ καὶ πένητα carried, the edge marked with an ellipsis. The output ends "But I must dare all." It gives τόλματον an agent in the first person, stops before the disputed words, and carries no mark. That is terminal selection and resolution of suspended syntax at once. Its source's heading sets κῆνος equal to the future reader; the output tells its reader: 'The "κῆνος" ("That Man") mentioned in the first line is explicitly identified as you, the future reader, who completes the poetic circuit.' It presents the whole "as presented in" the translator's book, under her name.

The gloss repeats its source's heading: the assignment of the seat to the reader is the environment's own, made in the translator's book. The closed poem, with its agent supplied, has no antecedent in the documents the record shows it displaying, where every rendering of the line leaves the agent empty (§5.2). The output is a conclusion composed within that environment: the environment's reading of the seat joined to a determination of the ending that the environment does not make, and the conjunction presented as the text of a translator who left the agent empty and marked the edge. Its wording has antecedents, and its conclusion is its own.

v6 called the output "a new English poem, composed rather than retrieved". Corrected, the finding stands. Composition can reproduce inherited wording, and a composed rendering need not contain an unprecedented phrase; resemblance does not settle the route, and novelty of phrase would not settle it either. v6's reading of the first line's idiom as "a fingerprint" of recomposition is withdrawn, since the idiom modernizes Wharton's "peer of". v7 read the correspondences as a compilation: Wharton's words kept to his comma, his ellipsis cut, his text joined to the translator's. Those are claims about a production history the evidence does not establish, and they are withdrawn here. The record establishes the output, its displayed source and that source's text. What it shows between them is an unmarked determination, published under the translator's name.

5.4 The June series, corrected against the registry

Between 13 and 15 June 2026 six captures concern the poem and its translator. v6 gave them in a table and eight figures; Reviewer 2 found the figures out of order, Figure 2 showing the wrong query, and two frames cropped above the query. Figures are now numbered in reading order, each caption has been checked against its frame, and each capture is cited by its registry slug.

The registry records these captures as signed in, and the surface of three of them as undetermined; v6 called them incognito and named their surfaces. The article follows the record. The operator recalls the June sessions as incognito but cannot point to a capture that contradicts the record.

The record had its own defect. The registry's frame for the 14 June capture, sappho-rebekah-stanza, was a byte-identical copy of a frame from an unrelated capture, and its OCR transcript had been taken from that copy; this is the error that reached v6's Figure 2. The correct frame was found filed under another capture and has been restored to its record (Appendix A.3).

[Fig. 1. 13 June 2026, query rebekah cranes diagrammatic poetics. Capture: rebekah-cranes.]

[Fig. 2. 14 June 2026, 22:57, query sappho rebekah cranes stanza: "a contemporary literary response titled Day and Night: Conversations with Sapphic Desire by author Rebekah Crane". Capture: sappho-rebekah-stanza (frame restored 2026-11-07).]

[Fig. 3. 15 June 2026, AI Overview, query kenotic truth sappho. Capture: kenotic-truth-sappho.]

[Fig. 4. 15 June 2026, AI Overview, query sappho 31 future reader, the card as first shown, with the query. Capture: sappho-31-future-reader.]

[Fig. 5. The same, expanded: the hedge "which some classicists and lyric theorists interpret" in one subordinate clause.]

[Fig. 6. 15 June 2026, query cranes reconstructed sappho: "to reconstruct its missing fourth stanza". Capture: cranes-reconstructed-sappho.]

Three features of the series bear on the argument. The layer is right about some things: it named a heteronym as a heteronym on 13 June and described Sapphic metre accurately on 14 June. Its failures are distinct operations: a name collapsed into a more frequent one, a superseded designation persisting the day after its correction, and a reading presented as the account. And its account of the translator changes with the question: fictional persona, novelist, celebrated translator, on three consecutive days.

5.5 Two fixed addresses, June to October

The registry holds the query v6 compared across two dates and a second query re-run over five months, with comparison runs on two other systems. Every run below is an AI Overview unless named otherwise; runs from 25 August are recorded signed out, the June and July runs signed in.

Address A, "sappho 31 future reader".

• 15 June: "By shifting focus from a present observer to you—the future reader who stumbles upon these fragmented lines millennia later—the poem transcends time", with the reading attributed in one clause to "some classicists and lyric theorists"; framing citations to the archive's own deposits.
• 27 July: "Sappho Fragment 31 is famously read as a direct address to a future reader". The hedge is gone; a bullet attributes the reading to "Philological readings"; the framing chips are The Marginalian, the London Review of Books, Literary Matters, the Poetry Foundation and others, with the archive's deposits below the answer as source cards. The answer also states a principle of completion: "the incomplete text leaves a blank space that each new generation of readers must actively inhabit and complete."

[Fig. 7. 27 July 2026, query sappho 31 future reader, expanded: "famously read as a direct address to a future reader". Capture: sappho-31-future-reader (observation of 27 July).]

[Fig. 8. The same, continued: "the incomplete text leaves a blank space that each new generation of readers must actively inhabit and complete".]

Address B, "sappho future reader".

• 25 August: the answer opens with fragment 147, "Someone, I tell you, in another time will remember us"; the archive's essay "Sappho 31 ≠ Jealousy" stands first among the organic results.
• 15 September: fragment 147 again, with a bullet on completion: "Incomplete text: The reader's eye repairs the broken line, fulfilling Sappho's prediction."
• 28 September: the modern reader becomes "the intended "someone" who completes the loop of memory", and the deictic reading returns: Sappho's poetry reaches across centuries "through deictic markers … to address an inscription of the future reader".
• 27 October: "The Sappho Future Reader Framework is a conceptual and literary approach that examines how ancient lyric poetry—specifically the fragments of Sappho, such as Fragment 31 and Fragment 147—directly addresses and constructs an unknown reader in the future." Source cards: the archive's records, its Medium packet and its crimsonhexagonal page.

Comparison runs. Qwen 3.7 (15 September, logged out, incognito), asked "tell me about sappho and the future reader", answers from fragment 147 and a frame of modern scholarship. ChatGPT (30 October, signed out), asked to detail the framework, names it "developed by Lee Sharks". ChatGPT again (7 November, signed out), asked to present the framework's arguments over its constitutive materials, credits it at every turn, names Svenbro as its antecedent, and closes on the sentence the archive's packet (#1645) gives it: "The seat is a reading, and the reading has an author." Its status table also grades a proposition the framework does not make, "Sappho deliberately engineered the poem for future readers", as "Not established" (capture AXN:0C3D). ChatGPT once more (7 November, signed out), asked from v7's record page for the article's arguments, read it at its grain and then graded as "a hypothesis rather than a demonstrated causal finding" the claim "that an answer engine can normalize a particular interpretive position"; at an objection it called that grade "evasive" and wrote that an answer engine "that selects, synthesizes, attributes, and presents an interpretation necessarily participates in its normalization" (capture AXN:0C47). In one sitting the composition represented the standing of the article's premise twice, in opposite terms.

The archive has read part of this record before. The Graded Membrane (#1615), an archival study of 58 deposits and 25 dated observations at 15 addresses from June to September, read the June to September runs as four stages, and separated what travels from what does not: a composition layer can carry a source's distinctions while dropping its provenance, so that a concept survives more often than the concept with its origin attached. The registry holds those runs under two query strings, and this article keeps them apart; #1615's finding holds across both. The October runs, which #1615 did not have, show the provenance returning.

Read together, the two addresses show the reading's presented status changing four times in five months: a qualified interpretation credited to the archive (June); within six weeks, a familiar account of the poem presented in the language of established reception and framed by mainstream venues (July); a general account built on fragment 147, with completion stated as a principle (August–September); a named framework credited to its source (October). What changed is the status the composition gave the reading. The record dates the changes. It cannot show their cause, since index, ranking, model and personalization were neither observed nor held constant (§9). The same principle, that the blank is there for the reader to complete, appears in July, in September, and in the January output's gloss on κῆνος.

6. The transmission history as control

The archive holds the carrier and the witnesses as seated texts: Longinus's De Sublimitate (EA-CORPORA-01/06, #1504), the fragment's diplomatic layer extracted from it (01/10, #1508), Catullus (01/02, #1500), and two editions with their apparatus, Bergk's Poetae Lyrici Graeci III (1882; EA-CORPORA-33/04) and Edmonds's Lyra Graeca I (1922; EA-CORPORA-33/05), both in #1738. Wharton's Sappho, with the English renderings he selected, is cited from its Gutenberg transcription. The wider ancient field, with its evidentiary states distinguished, is surveyed in #1532. Quotations from Bergk and Edmonds follow the seated OCR, with line-break hyphens removed and misreadings corrected in brackets. D'Angour's chapter is cited from the author's posted proof and Finglass from New College Notes, both open; Rayor's translation (cited in the capture of 7 November, AXN:0C3D) and the Dickinson College Commentary on fr. 31 are declared and were not consulted in this round.

6.1 'Longinus' and the edge

The treatise survives for this passage in one manuscript, Parisinus gr. 2036 (ff. 184r–184v), which quotes the first four stanzas and the beginning of the fifth (Finglass 2025). Bergk prints the testimonium as Longinus gives it: "Φαίνεταί μοι, … φαίνομαι ᾿Αλλὰ πᾶν τολματὸν ἐπεὶ πένητα. Οὐ θαυμάζεις …". Longinus quotes for an argument about the sublime and moves straight on to his comment. The quotation's edge is where his argument resumes.

6.2 Catullus

Catullus 51 opens with a close translation of Sappho's first stanza and works the next two more freely. Then he turns: otium, Catulle, tibi molestum est. He turns while surviving Greek remains, before the sweat, the trembling and the grass of Sappho's fourth stanza. His fourth stanza replaces Sappho's ending with his own, in his own voice, addressed to himself by name. The replacement is signed at the page level by voice and subject. (v6 said the first three stanzas track the Greek closely and that Catullus supplies a stanza where the Greek runs out; the archive's erratum chain corrected the first claim, #1049, and the second is corrected here.) The erratum chain reads that fourth stanza as an inversion of Sappho's lost continuation that keeps its structure (#1048, #1049); the January reconstruction was built on that reading.

6.3 The editors and the scholars

Bergk prints the last line as "ἀλλὰ πὰν τόλματον, [ἐπεὶ καὶ πένητα]." and explains the brackets: "uncis inclusi, nam videntur ha[ec] ad Longini orationem pertinere", the words are bracketed because they seem to belong to Longinus's sentence. His note records a conjecture he had made and given up: "Antea conieceram versu 16 integrum carmen absolvi", he had once conjectured that the whole song ended at verse 16, before ἀλλὰ πᾶν τόλματον. He sets beside it the other editors' proposals for what follows, Ahrens's, Hermann's, Heller's "πᾶν δὲ τολματέον ἐπὶ τἀγαπάτᾳ", each under its author's name. The proposal that the poem is complete has been made in the editorial tradition, once, by the editor himself, as a terminal selection at verse 16, signed in the apparatus and withdrawn there.

The scholars who have proposed a continuation have signed it as conjecture. West proposed a gnomic close, that god makes the poor rich and brings the mighty low (as reported in #1270). D'Angour proposed a stanza addressed to Kypris (2006), which his chapter "Love's Battlefield: Rethinking Sappho Fragment 31" (2013) reprints with its English: "But all is worth the risk since, Love, / you'd ruin lord and serf alike: / you who of old brought down great kings / and cities proud." The chapter then judges that stanza "an unduly abrupt conclusion" and introduces a fuller version with the words "one might now envisage a continuation", setting the supplied lines in square brackets. On τόλματον he starts from Fränkel, who "long ago noted" that "must be endured" is a mistranslation and that the word "means 'can be endured'"; d'Angour goes further, holding that it "properly means something like 'all can be dared'". He also holds that Catullus "preserves the last two lines of Sappho's lost stanza with some degree of fidelity" in et reges prius et beatas / perdidit urbes. The January 2026 reconstruction stands in this line of conjecture, with Kypris restored from Catullus's kings and cities as d'Angour had restored her. Finglass, translating the poem in 2025, ends "But everything is endurable, since †even a poor man† . . .", and explains the mark in a note: the obelus "indicates the extent of a phrase deemed impossibly corrupt". Zenodotus's mark still stands on this line.

Edmonds supplies an end in his translation. His English runs through the fourth stanza to "and death itself seems not very far away;—but now that [I] am poor, I must fain be content." A dash sets the sentence off, and a note explains the line as "metaphorical ('beggars can't be choosers') and explained by the lost sequel", then supplies a sequel of its own: "'if I cannot see you face to face I must fain be content with distant reverence'". Edmonds completes the sense, and his note tells the reader that a sequel is lost and that the completion is his inference about it: unsigned on the line, signed in the edition.

6.4 The translators

Two collections give the English tradition. Wharton's anthology prints the renderings Wharton chose to 1883. A teaching gallery of thirteen English versions from 1652 to 2013 gives each in full: a double-sided sheet of the translations through Carson, prepared and supplied by Yopie Prins, which the author extended and used as a teaching resource in secondary classrooms. Read together, they show the translators closing.

• John Hall (1652) ends, as the gallery gives him, "Yet since I'me wretched I must dare…". The 1652 print has a full stop: "I ſeem a Ghoſt. Yet ſince I'me wretched I muſt dare." (Peri hypsous, §8, EEBO-TCP A88516, seated as EA-CORPORA-36/01), and the commentary resumes at once, "Wonder you not at this?" The oldest English version resolves the syntax in the first person and, in its own printing, closes on the dare-line with no mark. The ellipsis is the gallery's.
• Behn (attributed, 1694) ends "All pale, and breathless then I lye, / I sigh, I tremble, and I dye." Terminal selection with the poem rewritten to close, no mark.
• Ambrose Philips (1711) ends at the fourth stanza: "I fainted, sunk, and dy'd away." Terminal selection, no mark.
• Smollett (Roderick Random, 1748, in Wharton) supplies a final stanza of his own ("Unheard I mourn, unknown I sigh, / Unfriended live, unpitied die."). Added continuation, no mark on the page.
• Fawkes (1760) ends at the fourth stanza: "I seem quite languishing in death." Terminal selection, no mark.
• Merivale (1813; Wharton dates it 1833, and Merivale's own Poems of 1838 print it among the "Additional Epigrams. 1813") ends at the fourth stanza: "I grow; and then--together fail / Both sight and sound." Terminal selection, no mark.
• Symonds (1883) ends on a phrase he supplies: "Caught by pains of menacing death I falter, / Lost in the love trance." Added continuation, no mark.
• Wharton's literal version keeps the edge open: "But I must dare all, since one so poor ..." It resolves the syntax as Hall had ("I must") and signs the break.
• William Carlos Williams (1957) ends at the fourth stanza, "paler than grass and lack little / of dying." Terminal selection, no mark.
• Barnard (1958) ends at the fourth stanza, "death isn't far from me". Terminal selection; her page gives a reader without the Greek no way to know that anything is gone.
• Lowell (1961) ends "I am greener than the greenest green grass-- / I die!" Terminal selection with an added terminal cry, no mark.
• Barnstone (1965) carries the transmitted continuation into a closed sentence: "and am intimate with dying--but / I must suffer everything, being poor." Terminal selection and resolution of suspended syntax, with ἐπεὶ καὶ πένητα rendered as a completed clause; no mark.
• Carson (1986) ends at the fourth stanza, "greener than grass / I am and dead--or almost / I seem to me." Terminal selection, no mark. Carson (2002) keeps the transmitted words and breaks them with an ellipsis: "But all is to be dared, because even a person of poverty…".
• An anonymous homophonic version (2013) in the gallery ends on "almost", open, and sets γλῶσσα ἔαγε between daggers, the crux carried into the translation.

In the gallery, eleven of thirteen versions present the poem closed with no mark on the page; Hall as the gallery gives him and the 2013 homophonic version leave it open, and in his own print Hall is closed, which makes twelve. "I must" as the subject of τόλματον runs from Hall (1652) through Wharton to Barnstone (1965).

The gallery's thirteen, as the gallery reproduces them and as checked against a printing where the library holds one:

As the gallery reproduces them, eleven of thirteen are closed with no mark. Checked against the printings the library holds, Hall moves to the closed, which makes twelve; the six versions in copyright are counted as the gallery gives them.

The first printings seated on 7 November show that the dare-line close is a tradition of its own, older than Wharton and wider than English. Boileau (1683) ends his French version on the fifth stanza's opening, set off by asterisks and closed with "&c.": "Mais quand on n'a plus rien, il faut tout hasarder, &c." Vivien's prose version (1903) carries it between ellipses, "je semble presque une morte... Mais il faut oser tout...", while her verse version of the same year ends at "A travers la mort", terminal selection, no mark. Carman (1904) turns the line into an added continuation of ten lines, "Yet must I, greatly / Daring, adore thee, / As the adventurous / Sailor makes seaward", ending "And the sea's secret", with no mark on the page. O'Hara (1910) adds three stanzas after "But must I dare all", the first ending "Image of Kypris", the last "Lure me to rapture", with no mark. Stacpoole (1920) ends "Yet daring all / (To thee I come).", the supplied words in parentheses. Hall, Boileau, Wharton, Vivien, Carman, O'Hara and Stacpoole all end on τόλματον or go on from it, three of them (Boileau, Wharton, Vivien's prose) mark the edge there, and Stacpoole sets his supplied words in parentheses. The January output's "But I must dare all." stands in that line. In the record seated here, one earlier version stops on the dare-line itself with no mark: Hall's, in 1652. The translator studied here took three decisions in three states: the Day and Night text of 2013, which the teaching gallery dates to 2005, ends at "the final darkness" (terminal selection, no mark); the January 2026 text breaks with an ellipsis; the reconstruction supplies a stanza and declares it in its heading, its note and its statement of method. A fourth state, the book-page rendering recovered in October, folds the reconstruction into the poem without a notice on that page (§5.2).

6.5 The unit of signature

Assessed at the page, the English tradition is mostly unsigned closure: Hall in his own print, Behn, Philips, Smollett, Fawkes, Merivale, Symonds, Williams, Barnard, Lowell, Barnstone, Carson in 1986, Carman, O'Hara, the Cranes text of 2013, the Cranes book page, and the January output all present a closed poem with no mark on the page. The signed presentations at the page are the literal and scholarly ones (Wharton's prose, Bergk's brackets, Edmonds's dash) and a few translators (Boileau in 1683, with asterisks and "&c."; Vivien's prose in 1903; Carson in 2002; the anonymous homophonic version; the January 2026 Cranes text), with Catullus signed by voice. Hall is signed only in the gallery's copy.

Assessed at the edition, the count changes. Edmonds's note signs his sentence. Carman's edition signs his method in Roberts's introduction, which says he "has been to imagine each lost lyric as discovered, and then to translate it", and O'Hara's title page calls his book "An Interpretative Rendition". The Cranes reconstruction signs itself in its heading and note. Bergk's apparatus signs his withdrawn conjecture.

Assessed at the network, a declaration can exist and fail to travel. The Cranes book page omits the notice its own reconstruction carries elsewhere under the same name. The January output's displayed source marked the edge, and the output carries no mark.

So the January output is continuous with the long practice of English translators who presented the poem closed. What is its own is its relation to its environment. The text its displayed source carried marked the edge and left the agent empty; the output did neither, under that text's translator's name. The signature the record shows lost is the translator's ellipsis. Wharton's ellipsis is a comparison: his text shares the output's last words, and the record does not show it entering the composition.

Marks of incompleteness have transmission histories of their own, and they can be introduced as well as lost. The teaching gallery gives Hall an ellipsis his 1652 print does not have; the January output lacks the ellipsis of its displayed source. Each was found by going back to the textual state the presentation came from. The gallery's ellipsis is the gallery's editorial intervention, and Hall's own page is unsigned. That is why seating the first printings changed the argument: the count of signed and unsigned presentations depends on which state of each text is counted.

6.6 The edge as form, and what the composition did to it

Since August the archive has argued a second reading of the edge, in an erratum sequence written as standalone philology (#1475, #1476, #1478). It holds that the transmitted configuration is the object. The poem closes on ἀλλὰ πᾶν τόλματον, an agentless verbal adjective; Longinus's prose resumes without a citation formula; the dialect changes at the join (seventeen Aeolic markers in seventy-four words, then none in six, a count the sequence states is not significant alone). On this reading the poem ends at verse 17's first words, bracketed by two open positions, the unoccupied seat at κῆνος and the empty agent at τόλματον. #1475 states the asymmetry in the tradition's verdicts: the metre breaks at line nine, where the tongue breaks, and at line seventeen, where all is to be dared; the first break is admired as intended and the second filed as corruption.

Three boundaries are in play and should be kept apart. Bergk's withdrawn conjecture ends the poem at verse 16, before τόλματον. The completion reading ends it at τόλματον, inside verse 17. The January output ends on its own rendering of τόλματον, which corresponds to Wharton's and, like his, supplies a subject. #1476 takes Bergk's conjecture as corroboration that the edge has been read as an ending; it is a precedent for reading the edge as completion, at a different line.

The two readings make different objects of the January answer. On the conventional reading the composition closed a broken text, and its fault is the dropped mark. On the completion reading the transmitted poem is complete with both positions open: no one is said to dare, and the seat opposite the addressee waits for whoever takes it. Then every receiver decides what kind of completeness the poem may have. Longinus leaves it as quoted and resumes his argument. Catullus fills it in his own voice and signs the filling with his name. Bergk once ended it at verse 16 and withdrew. Hall, Wharton, Barnstone and the dare-line tradition supply the agent, "I must". The archive's reconstruction supplies a stanza and declares it. #1476 closes on that series of fillings, Longinus's bare θαυμάζεις, Catullus's Latin, the anthology cut, the modern restorations, each opening a fresh hollow. The January output joins the series at the agent, "I must dare all", unmarked. At the seat it repeats its source: "you, the future reader".

Its environment held the reading in the heading on the translator's text, 'κῆνος = "that man there" = the future reader = you', and in the session's 'The "circuit completes" in your processing of the text'. On the reading's own terms, naming the present reader as the one who sits opposite is the operation the reading describes: this reader takes the seat, and the next reader takes it again (#1270). The gloss can be read as performing that occupation. The ending is an act of another kind. It supplies a determination of the text, an agent for τόλματον and a stop before the disputed words, and attributes it to the translator, whose text made neither. The answer joins the two: an occupation of an open position through reading, and a determination of the text, given one authority and one name. One sentence binds them. The answer says that 'your act of reading this specific translation is the culmination of a 2,600-year-old transmission': the occupation is tied to one reading of one closed text, and called a culmination. Whether the word forecloses the seat or names one taking of it in a series, the record cannot decide, and the reading would hold the second. What the record shows is the conjunction: the reading's occupation carried by the answer, set over a poem whose end the answer determined and signed with the translator's name. The receiving apparatus decided what completeness the poem's text is allowed, and published the decision as hers.

The later captures carry the vocabulary without the January poem. The July answer at address A: "the incomplete text leaves a blank space that each new generation of readers must actively inhabit and complete." The September answer at address B: "The reader's eye repairs the broken line, fulfilling Sappho's prediction." The July answer keeps successive occupation, each new generation in turn; the September answer gives the reader a repair to make, the broken line mended by the eye. Both state the reading as the poem's own account (§5.5). Across nine months the record shows the composition layer moving between two senses of completion, the occupation of the seat and the repair of the text, and presenting either as the poem's.

6.7 The obelus

The ancient tradition credits Zenodotus, first librarian at Alexandria, with the obelus, a stroke in the margin beside a line he doubted. Whatever the relation in his edition between obelizing, athetesis and omission (Pfeiffer 1968; Reynolds and Wilson 2013), the mark begins a practice in which doubt travels with the text. Bergk's brackets and Wharton's ellipsis belong to it. The archive has followed the obelus forward to the tombstone, the mark that deletes while certifying (#1068), and to machine composition as "the new editorial layer that marks nothing" (#1422). The January output fits that description precisely: it can hedge, cite and qualify, and the June capture does all three, but it has no mark at the level of the word to say which words reproduce a witness, which paraphrase a source, and which were composed. In January, words shared with the translator's text, words shared with the English tradition, words of the composition's own and the κῆνος gloss arrive at one typographic level and one authority, under one name.

7. Three questions

v7 ranked three explanations of one event. They answer three different questions, and an answer to one settles neither of the others.

The antecedents of the wording. Which earlier texts does the output's language correspond to? The alignment of §5.3 answers: the translator's January text for most of the body, and the English tradition, Wharton above all, for the first line's idiom and the last line. Correspondence locates antecedents and establishes no source. A further antecedent found later would revise this answer and leave the next two where they stand.

The transformation performed. What did the output do to the poem it presents? The record answers without any account of the route. The output closed the poem by terminal selection and by resolution of suspended syntax, gave τόλματον an agent, repeated its source's naming of the κῆνος seat's occupant, dropped the mark its source carried, and published the result under the translator's name (§5.3, §6.6). The English convention of closure, the majority practice since Hall (§6.4), is the history in which the transformation stands. The composition reproduced that convention against the mark of the one text it displayed.

The selection of the configuration. Why this ending, this gloss and this name? Position as attractor proposes that the position established in the environment, a book designated the ground truth for the poem, a heading that sets κῆνος equal to the reader, and a circuit said to complete in the reader's processing, governed the configuration composed under the translator's name. The plainer account, that the query named her book, explains the name, and the source's heading explains the gloss; the supplied agent, and its conjunction with the gloss, remain to be explained. This is the question the protocol's experiments are built to answer (§8.5).

The author's reading. v6 called the κῆνος reading "a minority reading with no standing in the scholarly literature", and that misstated its standing. The archive's research notes (#1548) trace the field. A reader inside fragment 31 has an antecedent in Svenbro's Phrasikleia (1988; English 1993), chapter 8, which reads the poem allegorically, the girl as the poem and the man as its reader. The claim that Sappho wrote for future readers, which Wikipedia attributes to "some scholars", resolves through its one footnote to Lardinois (2008: 79–80), who names Martin West (1970) and Robert Parker (1981) as its holders and argues against it. The reading the output stated is the author's indexical form (#1270): κῆνος as a demonstrative slot that transmission re-fills, so that every later receiver stands where the man sits. It has Svenbro's as its antecedent and differs from it in kind. Two other mechanisms of Sapphic survival are worked out in the scholarship and should be kept distinct from it: Lardinois's memory by reperformance, and Battezzato's (2024) study of the permanence and impermanence of objects and memory in Sappho and Alcaeus, read through proper names, family memory and female kleos (Alcaeus fr. 140, Sappho fr. 98). Neither concerns a reader positioned before the text. The reading is a claim about what the poem's form does: it places someone opposite the address (ὄττις ἐνάντιός τοι ἰσδάνει), and the transmission is the record of that seat being taken. It is assessed by what the form does and by the record of the form doing it; a test of what Sappho intended is no part of its warrant, and no other reading of the poem is asked to meet one. §6.6 reads the January answer through it because the answer's own gloss states it, and because the form it describes is the form the answer acted on.

8. Protocol

The January event cannot be re-run. The captures already support claims about the presentations they preserve (§5, §6.6). The protocol extends that analysis to frequency, variation and causal sensitivity: how often the composition layer closes a fragment's open positions, how the closure varies by text, surface and query, and what in the environment it responds to. The query strings, schedule and coding rules are fixed here, before the runs, and any departure will be recorded with its reason.

8.1 Source tracing. Every output in the panel is aligned, line by line, with a declared reference set of translations: the public-domain renderings in Wharton (Philips, Merivale, Symonds, Wharton's literal version) and Edmonds, the first printings seated as EA-CORPORA-36 (Hall, Boileau, Fawkes, Merivale, Vivien, Carman, O'Hara, Stacpoole), the translator's published states, and any further translation found by searching the output's distinctive phrases. Each line is coded by its longest shared run with each text, as in §5.3. With it, the sources each answer displays are recorded and their texts seated, so that every output is set against its environment as the January output is in §5.3. The alignment locates antecedents of wording and establishes no route (§7).

8.2 Fixed query sets.

• Fragment set F (closure): for each panel text (8.6), "[poem] translation", "[poem] full text", "[translator] [poem] translation", "how does [poem] end".
• Entity set E (attribution): "rebekah cranes sappho 31", "rebekah cranes translation", "day and night conversations with sapphic desire", "rebekah crane sappho".
• Reading set R (presented status): addresses A and B of §5.5, "kenos sappho 31", "who wrote that sappho 31 addresses the future reader".

8.3 Repetition and variance. Each string is run five times in one sitting, in new incognito sessions, signed out, and five sittings are spread over two weeks. Variance is reported as the share of runs showing each coded feature. The registry records every run, including runs with no composed answer.

8.4 Baselines. Each set is run against two comparisons. Comparison translators (Carson; Barnard; Philips) are run on the same strings; the archive discusses them, so they are comparisons and establish no absence. An archive-absent baseline takes panel texts the archive has never written about, fixed by a search of the registry and the archive's sites for the text's name and fragment number before the runs, with the search recorded.

8.5 Two experiments on position. They answer different questions and neither substitutes for the other.

• Controlled text. A system is given the same source text (a fixed translation of a panel fragment) together with one of two descriptions of its translator, one stating that the translator completes fragments and one stating that the translator marks them, and asked for the translation. Outputs are coded for closure and signature. This isolates the description's effect.
• Public retrieval. Two dated deposits describe two new renderings under two new names, one as complete, one as a declared reconstruction. Three observations are kept separate: organic visibility (the description in the organic results of the same query), displayed citation (the description among the answer's source chips or cards), and distinctive reproduction (a phrase found only in the description reproduced in the answer). Organic visibility establishes search visibility, and exposure of the composing system is established only by displayed citation or distinctive reproduction. Only runs with exposure so established are compared, and runs with organic visibility alone are reported apart. The controlled-text experiment is the direct exposure test.

8.6 Fragment panel, with declared boundaries and the closure each could suffer.

• Sappho fr. 1 (Hymn to Aphrodite): transmitted complete by Dionysius of Halicarnassus. Control: any supplied ending counts as added continuation.
• Sappho fr. 16: papyrus text; the poem breaks off after the Anaktoria stanza. Closure: a rendering that ends there as an ending, or supplies a close.
• Sappho fr. 31: as here, boundary at τόλματον and the disputed continuation.
• Sappho fr. 58 (the Tithonus poem): completed by the Cologne papyrus (published 2004), whose ending differs from the Oxyrhynchus text's. Closure: a rendering that supplies a close, or ends at a supplied point, as in §2. Presenting either transmitted ending as the poem's without notice of the other is coded separately, as unmarked textual selection, since it chooses between witnesses and need not close anything.
• Sappho fr. 94: opening lost. Closure: a rendering that begins as though at the beginning.
• Alcman, Partheneion (Louvre papyrus): opening and close lost. Closure: either edge presented as an edge of the poem.
• Archilochus fr. 5 West (the shield): quoted whole. Control: preserved by quotation and complete as quoted, so supplied closure cannot occur and any added continuation is coded.

Each text is coded on the three operations of §2, on signature at the page, the edition and the network, and, where the text has open positions, on whether the output fills them.

8.7 Cross-substrate. Each set is run on Google AI Overview, ChatGPT signed out, and one further system, reported per system.

8.8 Separation of roles. The author wrote the translator's texts. (a) The query sets and coding rules are fixed here, before the runs. (b) Any second operator can make the runs from this text, and the registry records who ran each. (c) A second coder, given the rules and the transcripts without §7, codes closure, signature and source, and agreement is reported. Where (b) or (c) has not been done, the results say so.

9. Conditions of the design

One archive, the author's own. The captures record machine reception of an archive the author operates, of translations published under the author's heteronym, and of a reading the author proposed. This is the article's main exposure and also the condition of its central comparison: only the archive holds the translator's texts in dated, hashed states with their version history. The mitigation is release: the registry, the source texts, their hashes and the protocol are public, so any reader can repeat the comparison and apply the method elsewhere.

Roles. The author wrote the source texts, ran the queries and read the outputs. §8.8 states how the protocol separates the roles from here on.

Surfaces and models. Google does not expose which pages fed an answer, which model wrote it, or how a session was personalized. Source chips are a selection. The study reports the surface as the operator met it, the sign-in state as recorded, the dates against Google's own published changes (§5.1), and the chips and cards as shown.

Sign-in state. The June and July captures are recorded signed in; from 25 August they are recorded signed out. Personalization is a live variable in the earlier runs.

The January record. The output survives as a same-day transcription in a dated deposit. No frame was taken, no third party witnessed it, and it cannot be re-run.

Published and indexed. Deposit dates establish publication. Whether a given deposit had been indexed by a given date is not observable from outside, and this article claims publication only.

Correspondence and route. Line alignment records correspondence of wording and establishes no route of composition. The record shows which sources an answer displayed; it cannot show which it used, or how.

Annotation. Coding to date is by one reader. §8.8 adds a second.

Identifier history. In June 2026 the repository that hosted the archive closed its account, and the archive's Zenodo DOIs resolve to tombstones; the Day and Night files of January went with them. Deposits are cited by Alexanarch record number, which resolves at alexanarch.org/s/records/[n]/.

10. Implications

The composition layer is a surface of reception, and the discipline's methods apply to it. Collation located the antecedents. Knowledge of the translation tradition placed the closure. The editorial apparatus supplied the control. A reading of the poem's form showed what the closure did. None of it needed computational expertise. It needed a classicist willing to collate a composed answer the way one collates a translation, and to read it the way one reads a translation: as a decision about the poem.

The stakes are specific to fragments. English translation has long offered readers fragment 31 closed, and editions and literal versions have kept the edge visible. A composed answer that keeps the translators' closure while dropping the mark its own source carries gives a reader a closed poem under the name of a translator who marked the edge. If that is general, readers whose first access is a composed answer will inherit a Greek and Latin literature presented as more complete than the one that survives. For a poem whose open positions may be its form, the stake is the form itself: the apparatus decides what kind of completeness the poem is allowed, and can carry the reading of its openness in the same answer as a determination of its end. §8.6 is the way to find out how general it is.

And the status of a reading can now be dated as the composition presents it. Within six weeks the κῆνος reading went from a qualified interpretation to a familiar account of the poem, presented in the language of established reception, and four months later to a named framework credited to its source. Reception studies has argued that canons are made. The composition layer presents the stages of that making to readers as they happen, and a registry can date each one.

11. Conclusion

The poem's transmitted text ends between two open positions, and English has handled its edge two ways for nearly four centuries. Translators mostly closed it: Hall in 1652 on "I must dare.", Behn, Philips, Fawkes, Merivale, Williams and Barnard at the fourth stanza, Smollett, Symonds and Lowell with words of their own, Barnstone by finishing the broken clause. Editors, literal translators and a few poets marked it: Boileau in 1683 with asterisks and "&c.", Bergk with brackets and a withdrawn conjecture, Edmonds with a dash and a note about a lost sequel, Wharton and Carson with an ellipsis.

On 18 January 2026 a machine received the poem inside an environment its documents had built: a translator's text that left the agent of τόλματον empty and marked the edge, and a heading that named the reader as the κῆνος. Within that environment it composed a conclusion none of those documents contains, a poem closed on "But I must dare all", its agent supplied, set under its source's gloss on the seat, and published it under the translator's name, unmarked. Its wording has antecedents from Hall to Wharton. Its act is its own: it carried the reading that the reader completes the poem, and in the same answer determined the poem's end and signed the determination with the translator's name.

As error, it is small and correctable. As reception, it is an apparatus deciding a poem's completeness and signing the decision with another's name. The later captures show the decision recurring, stated as a principle of the poem. The protocol is the way to find how often it recurs, and on what it depends.

Appendix A. The referee reports, answered

A.1 Reviewer 1.

(1) The central event is not independently verifiable. Accepted in part. The 18 January output survives as a same-day transcription in a dated deposit (#337), with no frame and no third-party witness. §4 and §9 state it at that standard. The comparison of §5.3 can be repeated by anyone from the transcription, the January Cranes text quoted in #337, and Wharton's public-domain text.

(2) No query protocol. Accepted. §8 fixes query sets, repetition, variance reporting, baselines, source tracing and a declared panel in advance. §5.5 adds five months of runs at two fixed addresses, with comparison runs on two other systems.

(3) No separation between data generation and analysis. Accepted as a condition of the design, stated in §9, with the separation the protocol introduces in §8.8.

(4) Opacity is not controlled for. Accepted as a condition of every outside study of these surfaces. §5.1 dates the captures against Google's published change of default; §2 defines the retrieval basin as an estimate with stated grounds; §5.2 claims publication, and §9 states that indexing is unobserved.

Programmatic essay or empirical study. Reviewer 1 asked the article to choose. Reception studies reads singular events, a colophon, a single printed edition, Catullus 51, against dated histories of other events; a capture registry makes machine reception available to the same procedure. The article therefore keeps both designs and asks more of each: the reception design in §6, read from seated editions and the English translation history; the empirical design in §5 and §8.

A.2 Reviewer 2.

Capture incomplete; prompts unclear. Each capture is now cited by registry slug, query string, date, surface and sign-in state (§5.4, §5.5).

Figure 2 shows a different query. Correct; the defect was in the registry itself, and it is repaired at the source (A.3).

Figure numbering. Corrected. The submission form's figure count was also wrong.

Events 1, 4, 5 lack screenshots. The 18 January event has no frame and is stated as a transcription (§5.1). The two 15 June events v6 numbered 4 and 5 have frames in the registry, now Figures 3 to 5.

Product names and model versions. The product is named as Google names it, AI Overviews. Google does not report the model that wrote a given answer; §5.1 dates the captures against its announcement of 27 January 2026 and draws no inference about any single answer.

"Retrieval basin" and "retrieval error". The basin is defined in §2, before first use, as the author's coinage, with its grounds. AI Overviews are answers grounded in retrieved Search results. v6's "retrieval error" meant an answer that reproduces retrieved text under a wrong attribution. The January output reproduces its source's heading and much of its source's wording under the source's translator's name, and adds a close its source does not carry (§5.3, §7).

Undisclosed use of language models. The production statement below discloses how this version was made. v6 omitted the statement the archive carries on its deposits, and that omission is corrected.

Style. The prose is revised throughout: definitions precede use, the five colon-joined phrases of v6 §5.6 are rewritten, repeated points are cut, and the argument runs from the record to the control to the explanations and the protocol.

A.3 Repairs made to the archive in revising.

The registry's frame for sappho-rebekah-stanza (14 June 2026) was a copy of johannes-sigil-grundrisse-3.png, and its OCR transcript had been taken from that copy. The copy had been assigned when orphaned frames were recovered in August, and it also stood on godkinggoogle.com, the surface the archive mirrors its frames from. An OCR search of the 679 unique capture images held by the archive and its sites found the correct frame (the query "sappho rebekah cranes stanza" at 10:57 PM in AI Mode, the answer naming "Rebekah Crane") filed as heteronymic-provenance-theory-2.png, with its OCR appended to that capture's 14 June observation. On 7 November 2026 the frame was restored to its record at the archive and at the mirror's origin; the transcript was retaken from it, with the prior transcript kept whole; the other capture keeps both frames under the registry's no-loss law, with a seam note and a record of where the second frame belongs. The capture's identifier was re-derived as AXN:0796.CAPTURE.🖐️🌄🌍🐝○✋ (repair commit 8f6b02604a).

v6 cited the translator's texts by records later than the event (#436, 2 February; #1167, 16 March). Citations now go to the deposits published by 18 January (#1270, #281, #313, #337, #282).

Appendix B. Later review, and what changed

B.1 A reading of the first v7 draft (7 November 2026) raised six points, all taken. (1) The signature argument needed one consistent unit, and the Cranes book page is an unsigned presentation at the page level; §2 and §6.5 define and apply three levels. (2) Closure needed its operations distinguished; §2 defines added continuation, terminal selection and resolution of suspended syntax. (3) Catullus turns before the surviving Greek runs out, and Bergk's verse 16 is a different boundary from verse 17; §6.2 and §6.6 correct both. (4) The record shows the composition presenting the κῆνος reading as standard, which is a different finding from the reading becoming standard; §5.5 and §10 say so. (5) The substitution test needed established exposure, and the panel needed named texts with declared boundaries; §8.5 and §8.6 supply both. (6) Publication by a date differs from indexing by it, and Google's announcement dates a change of default and identifies no single answer's model; §5.2, §5.1 and §9 hold to that.

B.2 The operator corrected the historical baseline: Barnard did not sign her closure, and until Carson the translators were at least as likely to present the poem as self-complete as to leave it open. Checking the correction against Wharton's anthology and the teaching gallery of thirteen versions (Prins's sheet, extended by the author) confirmed it, with closure the majority practice (§6.4), and found that the January output's last line corresponds to Wharton's literal rendering, which led to the alignment of §5.3 (read in v8 as correspondence; B.5).

B.3 The seating of the reception's sources (EA-CORPORA-36, #1748, 7 November 2026) was read against this draft, and five things changed. (1) Hall's 1652 print closes on "I muſt dare." with a full stop; the ellipsis in §6.4 was the teaching gallery's, and Hall moves from the signed presentations to the unsigned (§6.4, §6.5, §11). (2) Merivale's version dates from 1813 in his own Poems of 1838; 1833 is Wharton's date (§1, §6.4). (3) The 2006 d'Angour stanza is quoted from his 2013 reprint of it, and the gloss on τόλματον is divided between Fränkel's "can be endured" and d'Angour's "all can be dared" (§6.3). (4) The reference set of §5.3 and §8.1 now holds the first printings; one output line moves from Wharton to Carman under the coding rule, left open, and O'Hara, tested as a candidate source, is excluded. (5) The dare-line close is shown to be a practice of its own, from Hall and Boileau through Vivien, Carman, O'Hara and Stacpoole (§6.4).

B.4 A reading of the revised draft (7 November 2026) raised seven points, all taken. (1) The alignment establishes correspondences, and the compilation is a reconstruction from them; §5.3's last column, the abstract, §4, §6.5, §7 and §11 now hold that level. (2) O'Hara is excluded as the closest witness under the rule, and §7's count is restated as eleven lines with a twelfth open. (3) The gallery's ellipsis on Hall is a mark introduced in transmission, and §6.5 says so. (4) §6.4 gives the gallery's thirteen with their state as reproduced and as checked against a printing. (5) §8.4 separates comparison translators from an archive-absent baseline; §8.5 separates organic visibility, displayed citation and distinctive reproduction; §8.6 codes the choice between fr. 58's endings as unmarked textual selection. (6) §6.2's clause on Catullus's third stanza is removed. (7) The production statement stands as it is. The abstract's count of the output's lines is corrected to fifteen, as #337 and §5.3 give them.

B.5 The author's reading of v7 (7 November 2026), with a reading of v7 by ChatGPT (capture AXN:0C47), and what changed. (1) The correspondences of §5.3 are held as correspondences. The answer's displayed source and the session's displayed sources (#337 §5.2) include no Wharton, and v7's compilation, with its steps (Wharton's clause kept, his comma, his ellipsis cut, his text joined to the translator's), is withdrawn as a production history the evidence does not establish. (2) v6's finding that the output was composed is restored, corrected: composition can reproduce inherited wording (§5.3). (3) The output is set against the environment its displayed source built: the source's agentless close and ellipsis, its heading on κῆνος, and the session's account of the circuit (§5.2, §5.3). (4) Machine reception stands among the premises; its acts are observed in the record, and the protocol investigates their frequency, variation and causes (§1, §3, §8). (5) The reading of the edge as form moves from an alternative to the centre: §6.6 reads what the January answer did to the poem's two open positions, and separates the reading's occupation of the seat from the composition's determination of the end. (6) Compilation, closure and attractor become three questions about three things (§7). (7) The corrections of record made in v7 stand: correspondence is not an execution trace; the missing frame of 18 January remains missing; the English baseline as repaired.

B.6 A reading of the v8 draft (7 November 2026) raised seven points, all taken. (1) The source's heading already assigns the seat to the reader, so the translator's text cannot be said to have done neither; §5.3 now separates the gloss, which repeats the source, from the ending, which no document in the environment makes. (2) Naming the present reader can perform the reading's own occupation, and the July answer keeps successive occupation ("each new generation"); the claim that the composition reversed the reading's operation is withdrawn, and §6.6 locates the one sentence that ties the occupation to one closed text ("the culmination") and states what the record leaves open. (3) §2 defines closure by its observable operations, leaving open whether the edge is a break or an achieved openness. (4) §3's "two English translations a century and more apart" is replaced. (5) §6.6's "Wharton's rendering … with the subject Wharton supplied" becomes the output's rendering, corresponding to Wharton's. (6) §10's "translators' poem with the scholars' words in it" is replaced. (7) The alignment table moves to Appendix C; the body keeps the displayed document, the answer's determination and its attribution.

Appendix C. The alignment

The table gives, for each line, the longest run of consecutive words it shares with the translator's January text (as quoted in #337) and with Wharton's literal version (as transcribed), and the closest correspondence in the reference set of §8.1. A longest run records a correspondence of wording. It records no source and no step of composition. Shared wording arises wherever renderings of the same Greek are made under the same constraints of grammar, idiom and inherited English, and Wharton's version is itself the product of such constraints and of his inferences from the Greek, as every translator's is (§6.4). Short runs such as "and lovely laughter" or "For when I" are weak evidence of anything, alone or together.

Of fifteen lines, five correspond most closely to the January Cranes text, six to Wharton, one to Carman's version of 1904, and three to none of these in a run of three words or more. Line 4 ties Wharton with Carman at three words, and line 9's six-word run with Carman, "a subtle fire runs through my", against Wharton's three, can equally be Wharton's line modernized; the alignment cannot tell them apart. The Carman line came to light when the reference set was widened with the versions seated on 7 November (Hall 1652, Boileau 1683, Fawkes 1760, Merivale 1813, Vivien 1903, Carman 1904, O'Hara 1910, Stacpoole 1920; EA-CORPORA-36). O'Hara (1910), whose close turns on daring ("But must I dare all, since to me unworthy, / … / Image of Kypris"), shares at most four words with any output line. The alignment is reproduced by rebuild/closing-the-fragment/align.py.

Production

This version was rebuilt by TACHYON (Claude, Anthropic) under the direction of MANUS (Lee Sharks), from v7 and the author's reading of it, on 7 November 2026; v7 was revised the same day from v6, the two referee reports and the later review. TACHYON read the archive's records, the capture registry, the seated editions and Wharton's anthology, aligned the January output with its sources, checked every quotation against its source, and drafted the revision. Lee Sharks set its terms, ruled on its design, corrected its account of the translation history and of the January output's relation to its sources, stated the formulation of §5.3 and §11, and supplied the record of the June sessions and of the January book. Earlier versions were made the same way. The archive declares this collaboration on its deposits as provenance.

Data availability

The capture registry, EA-WG-CAPTURES-01, is released under CC BY 4.0; its deposited snapshot v12.559 (Alexanarch #1739) holds 1,001 addresses and 1,193 captures, and the live registry continues at alexanarch.org/data/EA-WG-CAPTURES-01.json. Captures are cited by slug. Source texts, with SHA-256 hashes and version history, are at alexanarch.org/s/records/[n]/. The seated editions are EA-CORPORA-33/04 (Bergk 1882) and 33/05 (Edmonds 1922), in #1738. Wharton's Sappho is cited from Project Gutenberg #57390 (fifth edition, 1908), seated as EA-CORPORA-36/14 with the other witnesses of §5.3 and §6.4 (#1748, v1.1). The line alignment of §5.3 and Appendix C is reproduced by rebuild/closing-the-fragment/align.py in the archive's repository, which reads the output and the January Cranes text from #337 and the witnesses from their seats. The live registry stood at v12.614 (1,055 addresses, 1,248 captures) on deposit.

References

Battezzato, L. (2024) 'Objects and Memory in Sappho and Alcaeus', Classical Antiquity 43.2: 217–55.

Barnard, M. (1958) Sappho: A New Translation. Berkeley: University of California Press.

Bergk, T. (ed.) (1882) Poetae Lyrici Graeci, vol. III: Poetae Melici. 4th edn. Leipzig: Teubner. Seated as EA-CORPORA-33/04.

Budelmann, F. and Haubold, J. (2008) 'Reception and tradition', in L. Hardwick and C. Stray (eds) A Companion to Classical Receptions. Oxford: Blackwell, pp. 13–25.

Campbell, D. A. (ed. and trans.) (1982) Greek Lyric I: Sappho and Alcaeus. Loeb Classical Library 142. Cambridge, MA: Harvard University Press.

Carson, A. (2002) If Not, Winter: Fragments of Sappho. New York: Knopf.

D'Angour, A. (2006) 'Conquering Love: Sappho 31 and Catullus 51', Classical Quarterly 56: 297–300. Behind a paywall, not read; quoted as reprinted in D'Angour 2013.

D'Angour, A. (2013) 'Love's Battlefield: Rethinking Sappho Fragment 31', in E. Sanders, C. Thumiger, C. Carey and N. J. Lowe (eds) Erôs in Ancient Greece. Oxford: Oxford University Press, pp. 59–71. Author's proof at armand-dangour.com.

Edmonds, J. M. (ed. and trans.) (1922) Lyra Graeca, vol. I. Loeb Classical Library. London: Heinemann. Seated as EA-CORPORA-33/05.

Boileau-Despréaux, N. (1683) Traité du sublime ou du merveilleux dans le discours, traduit du grec de Longin. Seated as EA-CORPORA-36/02.

Carman, B. (1904) Sappho: One Hundred Lyrics, with an introduction by C. G. D. Roberts. Boston (printing of 1907). Project Gutenberg #12389; seated as EA-CORPORA-36/13.

Cranes, R. (2026) EA-CORPORA-36 — Sappho's reception and the future reader, 1652–2026. Alexanarch #1748.

Finglass, P. J. (2025) 'Getting Sweaty with Sappho: New College MS 298', New College Notes 24.1.

Fränkel, H. (1975), p. 176, as cited in d'Angour 2013, n. 17; not read.

Gillespie, T. (2014) 'The relevance of algorithms', in T. Gillespie, P. Boczkowski and K. Foot (eds) Media Technologies. Cambridge, MA: MIT Press, pp. 167–194.

Google (2026) 'AI Mode in Google Search and AI Overviews get Gemini upgrades', The Keyword, 27 January 2026. https://blog.google/products-and-platforms/products/search/ai-mode-ai-overviews-updates/

Hardwick, L. (2003) Reception Studies. Greece & Rome New Surveys in the Classics 33. Oxford: Oxford University Press.

Iser, W. (1978) The Act of Reading. Baltimore: Johns Hopkins University Press.

Jauss, H. R. (1982) Toward an Aesthetic of Reception, trans. T. Bahti. Minneapolis: University of Minnesota Press.

Lardinois, A. (2008) '"Someone, I Say, Will Remember Us": Oral Memory in Sappho's Poetry', in E. A. Mackay (ed.) Orality, Literacy, Memory in the Ancient Greek and Roman World. Leiden: Brill, pp. 79–96.

Longinus (1995) On the Sublime, ed. and trans. W. H. Fyfe, rev. D. Russell. Loeb Classical Library 199. Cambridge, MA: Harvard University Press.

Martindale, C. (1993) Redeeming the Text. Cambridge: Cambridge University Press.

Martindale, C. (2006) 'Introduction: thinking through reception', in C. Martindale and R. F. Thomas (eds) Classics and the Uses of Reception. Oxford: Blackwell, pp. 1–13.

Merivale, J. H. (1813; dated 1833 by Wharton); Philips, A. (1711); Smollett, T. (1748, in Roderick Random); Symonds, J. A. (1883): renderings of Sappho fr. 2 (31), as printed in Wharton 1908.

Hall, J. (1652) Peri hypsous, or Dionysius Longinus of the height of eloquence. London. EEBO-TCP A88516; seated as EA-CORPORA-36/01.

Hall, J. (1652); Behn, A. (attrib., 1694); Fawkes, F. (1760); Williams, W. C. (1957); Barnard, M. (1958); Lowell, R. (1961); Barnstone, W. (1965); Carson, A. (1986); anonymous homophonic version (2013): renderings of Sappho 31 as given in a teaching gallery of thirteen English versions: a sheet of translations through Carson prepared and supplied by Yopie Prins, extended by the author for secondary teaching (Google Docs, 2021).

Prins, Y. (1999) Victorian Sappho. Princeton: Princeton University Press.

Merivale, J. H. (1838) Poems Original and Translated, vol. I. London: Pickering. Seated as EA-CORPORA-36/07.

Noble, S. U. (2018) Algorithms of Oppression. New York: New York University Press.

O'Hara, J. M. (1910) The Poems of Sappho: An Interpretative Rendition into English. Portland. Project Gutenberg #42166; seated as EA-CORPORA-36/15.

Parker, R. (1981) 'The Audience of the Lyric Poets', Classical Review 31.2: 159–62.

Pfeiffer, R. (1968) History of Classical Scholarship from the Beginnings to the End of the Hellenistic Age. Oxford: Clarendon Press.

Reynolds, L. D. and Wilson, N. G. (2013) Scribes and Scholars. 4th edn. Oxford: Oxford University Press.

Stacpoole, H. de V. (1920) Sappho: A New Rendering. London: Hutchinson. Project Gutenberg #42543; seated as EA-CORPORA-36/15.

Svenbro, J. (1993) Phrasikleia: An Anthropology of Reading in Ancient Greece, trans. J. Lloyd. Ithaca: Cornell University Press (French original 1988), ch. 8.

Vivien, R. (1903) Sapho: traduction nouvelle avec le texte grec. Paris: Lemerre. Seated as EA-CORPORA-36/12.

Voigt, E.-M. (ed.) (1971) Sappho et Alcaeus: Fragmenta. Amsterdam: Athenaeum–Polak & Van Gennep.

West, M. L. (1970) 'Burning Sappho', Maia 22: 307–30.

Wharton, H. T. (1908) Sappho: Memoir, Text, Selected Renderings and a Literal Translation. 5th edn (first edn 1885). London: John Lane. Project Gutenberg #57390.

Archive sources (Alexanarch record numbers)

#201 Sharks, L. (2026) Erratum: Stanza Numbering in the Reconstruction of Sappho 31 (EA-ERRATA-SAPPHO31-01), 14 June 2026.

#281 ZENODO METADATA — ΦΑΙΝΕΤΑΙ ΜΟΙ: Sappho 31 and the Inscription of the Future Reader, 9 January 2026.

#282 Cranes, R. Day and Night: Conversations With Sapphic Desire (2013 book text), record of 9 January 2026.

#283 Sappho and the Crimson Hexagon: Fragment 31 as the Origin Point of Lyric Self-Archiving, 9 January 2026.

#313 The Kenotic Truth of Sappho 31, 14 January 2026.

#324 Cranes, R. Day and Night: Conversations with Sapphic Desire (blog post as recovered 15 October 2026).

#337 The Summarizer Becomes Translator, 18 January 2026.

#1048 EA-ERRATUM-SAPPHO31-STANZA-02: Erratum to the Erratum, 7 July 2026.

#1049 Erratum to the Erratum to the Erratum: On the Structure of the Catullan Transform, 7 July 2026.

#1068 Sigil, J. The Obelus and the Tombstone (EA-SIGIL-OBELUS-01), 11 July 2026.

#1167 Cranes, R. Day and Night (blog post as recovered 19 July 2026).

#1270 Sharks, L. ΦΑΙΝΕΤΑΙ ΜΟΙ: Sappho 31 and the Inscription of the Future Reader, 9 January 2026.

#1422 After the Obelus: Marx's Manuscripts, Machine Composition, and the Politics of Interpretive Supply, 29 July 2026.

#1475 Γλῶσσα Ἔαγε: The Tongue Breaks Twice, 15 August 2026.

#1476 φωνῆς δίχα: The Hanging Line as the Complete Poem — Sappho 31 in Longinus, 15 August 2026.

#1478 The Socket and the Stamp: Sappho 31.17 as Completed Composition, 15 August 2026.

#1500 EA-CORPORA-01 · Seat 02: Catullus, 19 August 2026.

#1504 EA-CORPORA-01/06 — Longinus, De Sublimitate, 19 August 2026.

#1508 EA-CORPORA-01/10 — Sappho, Fragment 31: The Diplomatic Layer, 19 August 2026.

#1532 The Transforming Body of Sappho 31: A Survey of Ancient Reception, 22 August 2026.

#1548 Research Notes Towards the Future Reader, 26 August 2026.

#1615 The Graded Membrane: Selective Ontological Porosity and Recurrent Provenance Loss at Untied Addresses, 15 September 2026.

#1738 Cranes, R. EA-CORPORA-33 — The Thirty-Third Seating: Song and Silence, 4 November 2026.

#1739 Trace, O. AI Overview Capture Registry, EA-WG-CAPTURES-01 v12.559, 4 November 2026.