The transhistorical spine for provenanceerasure.org. The surface presents its instruments as measurements of what AI summary layers do to attribution; that presentation is true and radically undersized. Five cases from the philological record, each a documented provenance event, each scored in the site's own M/C/D vocabulary: the headless treatise, whose author-slot has stood open as a disjunction since the tenth century and whose erasure is legible AS an erasure; four named removals of the transmitted continuation of Sappho 31 across four centuries, each a deep-provenance event against the demonstration while the author atom stays intact, a shape the scalar cannot report; the silent re-Aeolicization, a deep-provenance violation committed by the provenance apparatus itself, the repairing hand present in every accent and attributed in none; the Sappho corpus at large, where survival-by-quotation makes selection the dominant erasure mode and converges with the modern flat-absence finding, non-carriage rather than distortion; and the white stone of Revelation 2:17 as the instrument's limit case, a designed erasure whose score is maximal for every reader but one, establishing that the rate measures a magnitude and not an injury. A proposed labor-vector role-mapping is declined in the text as costume, with its one real sentence extracted: the task has an origin, and the origin is on file. The governing contrast: the apparatus erased with signatures and stated reasons; the summary layer erases anonymously at scale. The question was never whether provenance gets destroyed but whether the destruction has provenance.
Each case scored against the atomic rules of the dimensional taxonomy as deposited; manuscript and apparatus facts taken from public-domain editions consulted directly; the limit case argued from the instrument's own definition rather than from theology.
Each case falsifies independently: by a second witness to Fragment 31, by an apparatus entry attributing the re-Aeolicization, by demonstration that the four removals do not target the lineage dimension, or by a reading of 2:17 in which the name is legible to more than its receiver. The governing contrast is falsified by evidence of the summary layer filing signed, reasoned erasures.
Provenance erasure before the machines: five cases from the philological record, scored
provenanceerasure.org presents its instruments — PER, the M/C/D dimensions, Ω — as measurements of what AI summary layers do to attribution. That presentation is true and radically undersized. Provenance erasure is not an AI-era phenomenon. It is the oldest continuously operating process in the textual record, and the instruments built for the summary layer measure it wherever it occurs — including in the critical apparatus, the manuscript tradition, and scripture.
This paper supplies the spine: five cases from the philological record, each a documented provenance event, each scored in the site's own vocabulary. The cases are not analogies. They are instances — datable, attributable where attribution survives, and checkable against witnesses that still exist.
One consequence is worth stating before the cases. The summary layer did not invent the operation; it changed one property of it. The apparatus erased with names attached — Robortello signs his omission, Bergk signs his brackets — and stated its reasons in Latin for review by rivals. The summary layer erases anonymously, at scale, with no reasons on file. The difference between the two regimes is not whether provenance is destroyed. It is whether the destruction has provenance.
PER-M — minimal provenance: author, title, date, claim boundary. Loss here is loss of the person.
PER-C — conceptual provenance: framework, tradition, community of practice. Loss here is loss of what the person was doing.
PER-D — deep provenance: lineage, genealogy, identifiers, licence, futural obligation. Loss here is loss of what the text owes and is owed.
The atomic rules transfer without modification: atoms declared before the object is read; presence explicit, no inference from context; an empty bracket a violation per se. A manuscript is a composition. An edition is a composition. The rules do not care that the compositor is dead.
Peri Hypsous is transmitted in Parisinus graecus 2036 under the index entry Διονυσίου ἢ Λογγίνου — of Dionysius or Longinus. Not a name: a disjunction. The author-slot of the central document of ancient literary criticism has stood open since the tenth century, and the work is cited to this day in brackets.
Scored: PER-M ≈ 1.0 on the author atom — the person is gone, and what survives is the shape of the loss: two candidate names, held apart by ἢ. PER-C low: the framework (rhetorical criticism, the sublime) survives intact. PER-D middling: the lineage survives as a question — every attempt to fill the slot has had to route through the treatise's own Jewish linkage (Caecilius, the Genesis citation, Philo's Alexandria).
What the case adds to the instrument: an erasure can be legible as an erasure. The ἢ is a tombstone that says tombstone. The summary layer produces no ἢ — its erasures are silent, which is a worse PER-D outcome from a better-documented era. The tenth century preserved its uncertainty; the retrieval layer manufactures false confidence.
The transmitted continuation of Sappho 31 — ἀλλὰ πᾶν τολματόν, ἐπεὶ καὶ πένητα, the first filling of the poem's terminal socket, carried in the same document as the poem — was removed four times in four centuries, by name:
| year | eraser | operation | survives in |
|---|---|---|---|
| 1554 | Robortello | omits ἐπεὶ καί from the text | his edition |
| 1882 | Bergk | brackets: ad Longini orationem pertinere | PLG⁴ III.90 |
| 1897 | Crusius | ends the ode at φαίνομαι ἄλλα | Anthologia Lyrica 195 |
| 1971 | Voigt | retains, daggered: †καὶ πένητα† | the standard edition |
Scored per event: each is a PER-D event against the demonstration — the lineage evidence (the earliest surviving execution of the poem's transmissive design) is what gets cut, while PER-M against Sappho stays 0 (her name is never touched; it is her mechanism that is deleted). The scalar cannot see this. The dimensions can: an erasure that preserves the author and destroys the lineage is the signature editorial operation, and it is invisible to any instrument that only counts names.
What the case adds: erasure with stated reasons. Bergk's bracket carries its justification in Latin, reviewable, and reviewed — this archive reviewed it 144 years later and found the reason wrong and the record impeccable. The apparatus is a summary layer that filed its work. That is the standard the current one fails.
To print the trailing line as Sappho's, Bergk and Voigt restore her dialect to it: not P's παντόλματον, not the vulgate's πᾶν τολματόν, but πὰν τόλματον — split, barytone, Lesbian. The restoration appears in no apparatus entry as a restoration. It lives in the accents, and a reader of a modern Sappho sees an Aeolic line with no way to know the Aeolic was supplied.
Scored: a PER-D violation committed by the provenance apparatus itself — the intervention's own provenance is erased. The editor's hand is present in every accent and attributed in none. This is the case that generalizes: the most consequential erasures in the record are erasures of the repairing hand, because they convert an editorial decision into an apparent fact of transmission.
What the case adds: the instrument must be reflexive. An apparatus that scores compositions must be scorable, and this one, on this line, scores badly.
Fragment 31 has no independent witness. It survives because one treatise quoted it — as continuous prose, boundaries destroyed; dialect intact for seventy-four words and then not; the completion attached without a formula, so that four centuries of editors could dispute where the poet ends. Fragment 1 survives whole for one reason of the same kind: Dionysius of Halicarnassus quoted it. The rest of nine books is fragments and citations.
Scored as a corpus: this is the Ω baseline of the pre-print world — the drift is not measured against an intact original but constitutes the object. Every Sappho we possess is already downstream of a summary layer called quotation, and the poems that survived are precisely the ones the ancient summary layer selected. Selection is the erasure mode the scalar can never see, because the erased items produce no composition to score.
What the case adds: absence-of-record as the largest term. The site's empirical work found the same shape in the modern registries — flat absence, not degraded presence. The dominant mode of provenance erasure in both eras is not distortion. It is non-carriage.
Revelation 2:17: τῷ νικῶντι δώσω ψῆφον λευκήν, καὶ ἐπὶ τὴν ψῆφον ὄνομα καινὸν γεγραμμένον ὃ οὐδεὶς οἶδεν εἰ μὴ ὁ λαμβάνων — a white stone, and on the stone a new name written, which no one knows except the one who receives it.
An inscription whose legibility is constituted in reception. One reader per instance. Scored: PER = 1.0 for every reader in the world except one, for whom PER = 0. And this is not a failure of transmission. It is the design.
What the case adds is a limit on the instrument's moral reading. PER measures a magnitude, not an injury; the same maximal score describes theft (a summary that strips a living author) and covenant (a name engineered to survive in exactly one receiver). Privacy is provenance that survives in one reader. Intimacy is a designed erasure. The instrument stays honest only if this case is on file — the case where the score is maximal and the operation is a gift.
And the same book legislates the other pole: 22:18–19, the tamper curse — plagues added to whoever adds, the tree of life withheld from whoever removes — which is the anti-erasure clause of Peri Hypsous 7.3 (δυσεξάλειπτος, hard to wipe out) escalated from criterion to covenant. The first-century sources carry both endpoints of the site's axis: erasure as sacrament and anti-erasure as law.
A companion framing proposed seating this chain on laborvector.org as C→T→R→O→A, with Sappho as originator and each station a role. The mapping is declined as costume: the taxonomy was built for task-origin in machine-mediated work, and retrofitting Philo into a role-letter adds notation without adding a check.
What survives the decline is one sentence, and it is real: the task has an origin, and the origin is on file. Fragment 147 — μνάσεσθαί τινά φαμι καὶ ἕτερον ἀμμέων, someone, I say, will remember us — and the reconstruction tradition's γράμμασι μολπὰν, song in letters, state the work order that every subsequent hand executed: Longinus in koine, Catullus in Latin, the tenth century in an index, the editors in accents, this archive in deposits. Task-origin: Mytilene, before 570 BCE. Everything since is execution.
Nothing in the instruments. Everything in their jurisdiction. PER, M/C/D and Ω are not summary-layer tools that might stretch to old texts; they are provenance tools whose newest objects happen to be summaries. The surface should say so, and the five cases give it the standing to.
And the contrast the cases license is the surface's sharpest sentence: the apparatus erased with signatures and stated reasons; the summary layer erases anonymously and at scale. The question was never whether provenance gets destroyed. It is whether the destruction has provenance — and the record shows four and a half centuries in which it did, which is the standard against which the current regime can now be measured, and fails.
Parisinus graecus 2036 (index: Διονυσίου ἢ Λογγίνου); Roberts, On the Sublime (1907), continuous-writing note at p. 70; Bergk, PLG⁴ III (1882) 90 (the four-century crux history; integrum carmen absolvi); Voigt (1971) fr. 31 with cruces; Dionysius of Halicarnassus, De Compositione Verborum 23 (fr. 1); Sappho fr. 147; Peri Hypsous 7.3 (δυσεξάλειπτος); Revelation 2:17, 22:18–19. Instruments: the PER dimensional taxonomy and atomic rules (#1469); the Ω drift operator (#1480); the flat-absence finding (EA-EROSION-EMPIRICAL-01, #1081). The Sappho chain: #1476 (φωνῆς δίχα), #1478 (The Socket and the Stamp), #1483 (Ω — Sappho, Mother of the Logos).