When Vesuvius erupted in AD 79, it buried a library. The scrolls in a villa at Herculaneum were not burned to ash; they were carbonized into black lumps that look like charcoal — and some, when they were found in the 1750s, were thrown away as charcoal. The rest kept what the project reading them calls “a cruel bargain: its scrolls survived the eruption of Mount Vesuvius, but only by becoming too fragile to open. To read one was to destroy it.” In June 2026 a team announced that one of them, PHerc. 1667, had been read from beginning to end without being opened: scanned with X-rays, unrolled inside a computer, and searched for ink by a machine-learning model, with every reading checked by papyrologists. It turned out to be a work of Stoic ethics. The scroll that was read is itself a survivor of damage: two attempts to open it by hand, in the nineteenth century and in 1969, left only its inner core. And most of the team that read it began not as specialists but as contestants in an open challenge launched in 2023, the year a 21-year-old student read the first word from inside a closed scroll: “purple.”
In 1752 workmen of the Bourbon court came upon a villa at Herculaneum and, inside it, more than 1,800 scrolls of papyrus.[5] They are, in the 2026 paper’s words, “the only surviving large-scale library from classical antiquity.” Carbonized by Vesuvius in AD 79, the rolls “survived in a form that made reading and preservation fundamentally opposed: physical opening could reveal text, yet also irreversibly damage the material.”[1] Early handlers did not know what they had; an old account preserved in the reference record says many “had been thrown away as mere charcoal.”[5]
The method that changed this has three steps, and the companion page lets a reader try each one.[6] First, X-rays: the scroll is scanned at a particle accelerator in Grenoble, producing a three-dimensional image of its inside. Second, virtual unrolling: software traces the spiral of the papyrus through the scan and flattens it into a sheet. Third, ink: the ancient ink was carbon-based, almost indistinguishable from charred papyrus, so a machine-learning model is trained to find its faint traces. Papyrologists then read and check what it finds.[1]
The open Vesuvius Challenge, launched in 2023, turned these steps into public prizes. That year Luke Farritor, “the 21-year-old college student and SpaceX intern from Nebraska,” became “the first person in history to read an entire word from the inside of a Herculaneum scroll” — ΠΟΡΦΥΡΑϹ, “purple.” The Grand Prize of $700,000 went to Youssef Nader, Farritor and Julian Schilliger for “more than 2000 characters,” after a scholarly review whose results were “clear and unanimous” — while “roughly 95% of the scroll remains to be read.”[2]
On June 25, 2026, the project announced the next step: “We read an entire scroll — without ever opening it.”[3] PHerc. 1667 is 22 columns of writing over about 860 square centimetres. “The evidence points to a Stoic work”: a treatise on ethics whose final column names Aristocreon, “nephew and disciple of the great Stoic Chrysippus.”[1][3] In a second scroll, the team recovered a title — Philodemus, On Gods, Book 8 — which, the paper notes, “informs us for the first time that Philodemus’ work On Gods comprised at least eight books.”[1]
About 860 cm² of writing surface in PHerc. 1667, transcribed and checked by papyrologists; 33 cm² too fragmentary to read.[1]
A model that “finds” letters invites an obvious worry: that it is inventing them. The project faced it head on. In 2023 the organisers wrote that “machine learning models are infamous for ‘hallucinating’” and set out checks, including running each submission’s code independently.[2] The 2026 model is built so it cannot draw on what words are likely: each prediction sees only a small local patch, “making full-letterform hallucinations from linguistic or word-level priors practically impossible.”[1]
Then a physical check. For a different scroll, PHerc. Paris 4, an improved scan made the ink directly visible in the X-ray volume, with no model needed. That ink “coincide[s] with the Vesuvius Challenge Grand Prize Banner region” — the 2023 reading — giving “an independent physical validation case.”[1] The machine’s earlier guess and the later direct observation agree. Every reading of PHerc. 1667 was also reviewed and transcribed by papyrologists, who are among the paper’s authors.[1]
There is an irony in which scroll came first. PHerc. 1667 is only a fragment — 8 centimetres of a usual 19 to 24 in height — because people tried to open it. A nineteenth-century attempt left “the compact inner core,” and a conservator tried again in November 1969, revealing “no visible traces of writing” while “substantially reducing the dimensions of the surviving portion.”[1] The first complete reading was made possible by technology and preserved, in part, by the failure of earlier hands. And the people who did it mostly started outside the field: “Most of the Vesuvius Challenge research team first arrived as contestants.”[3]
More than 1,800 scrolls; some thrown away as charcoal, many destroyed in early opening attempts.[5]
DiscoveryThe outer layers are lost; a 1969 attempt finds no visible writing.[1]
DamageA student reads a word from inside a closed scroll; the Grand Prize follows for 2,000+ characters.[2]
The ChallengeTo read one was to destroy it. — Vesuvius Challenge, June 25, 2026 [3]
| Dimension | Evidence |
|---|---|
| Operational (D6) Origin · 84 | Launched in 2023 with public prizes and data; “most of the Vesuvius Challenge research team first arrived as contestants.”[2][3]An Open Contest Became a Lab |
| Quality (D5) L1 · 80 | Local-patch predictions make “full-letterform hallucinations … practically impossible”; Paris 4 ink later seen directly in the scan.[1]Built Not to Invent Letters |
| Employee (D2) L1 · 78 | A 21-year-old read the first word; the Grand Prize review was “clear and unanimous”; papyrologists transcribe and verify.[1][2]Students First, Specialists Checking |
| Regulatory (D4) L2 · 70 | Earlier opening attempts reduced PHerc. 1667 to an 8 cm core; reading without opening protects the rest of the library.[1][3]To Read One Was to Destroy It |
| Customer (D1) L2 · 60 | A Stoic treatise on ethics and a new book of Philodemus’ On Gods, with open data and code for anyone to inspect.[1][3]Closed for Two Thousand Years |
The cascade originates in D6 (Operational): an open competition that published scans and paid for progress, then hired the people who made it.[2][3] It propagates to D2 (Employee), where contestants became the research team and papyrologists review every reading,[1][3] and to D5 (Quality), where the model is designed against invented letters and one result is confirmed by directly visible ink.[1] It lands in D4 (Regulatory), the preservation of a library that cannot be opened without loss,[1][5] and in D1 (Customer), readers who can now see texts that have been closed for two thousand years. D3 (Revenue) is not scored.
This is the second case in the library’s archaeology thread. In UC-335, AI narrowed a desert and people walked to every candidate. Here the confirmation is a papyrologist’s eye and, once, the ink itself becoming visible. In both, the machine extends what people can examine; it does not replace the examining. The honest limit: the 2026 result is a preprint not yet peer-reviewed; the Stoic attribution and dating are the team’s working interpretation; and no outside papyrologist has yet commented on the record.
-- UC-336: How to Read a Book You Can't Open: 6D Amplifying Cascade
-- Herculaneum, AD 79; Villa of the Papyri found 1752, more than 1,800 scrolls; some thrown away as charcoal. Vesuvius Challenge launched 2023: first word PORPHYRAS (purple) by Luke Farritor; Grand Prize $700,000 to Nader, Farritor, Schilliger for 2,000+ characters, review clear and unanimous, ~95% of scroll unread. Angelotti et al., arXiv 2606.29085 (Jun 27 2026, preprint, 46 pp, 27 authors): PHerc. 1667 completely unwrapped and read under papyrological review; 22 columns, ~860 cm2; Stoic ethics naming Aristocreon; dated 2nd century BC; PHerc. 1667 is 8 of 19-24 cm height after 19th-century and 1969 opening attempts; PHerc. Paris 4 ink directly visible in scan coincides with 2023 Grand Prize region (independent physical validation); PHerc. 139 = Philodemus, On Gods, Book 8; local-patch model makes letterform hallucination practically impossible. Most of team first arrived as contestants.
FORAGE how_to_read_a_book_you_cant_open
WHERE text_unreadable_without_destroying_object = true
AND open_competition_produced_team = true
AND readings_reviewed_by_specialists = true
ACROSS D6, D2, D5, D4, D1
DEPTH 3
SURFACE how_to_read_a_book_you_cant_open
DIVE INTO read_without_opening
WHEN independent_physical_validation = true
AND peer_review_complete = false
TRACE open_science_reading_cascade
EMIT how_to_read_a_book_you_cant_open_signal
DRIFT how_to_read_a_book_you_cant_open
METHODOLOGY 88
PERFORMANCE 42
FETCH how_to_read_a_book_you_cant_open
THRESHOLD 1000
ON MONITOR CHIRP high 'A carbonized Herculaneum scroll, reduced to its core by two failed attempts to open it, has been read end to end without opening: X-ray scans, virtual unrolling and an ink-finding model built to avoid inventing letters, with papyrologists reviewing every reading. Most of the team began as contestants in a 2023 open challenge. The result is a preprint.'
SURFACE analysis AS json
Runtime: @stratiqx/cal-runtime · Spec: cal.semanticintent.dev · DOI: 10.5281/zenodo.18905193
For two thousand years, reading a Herculaneum scroll meant destroying it. Scanning, unrolling and ink-finding remove the trade-off.[3]
The first word was read by a 21-year-old student; most of the team that read a whole scroll began as contestants.[2][3]
The model sees only small patches, so it cannot borrow likely words — and in one scroll the ink later became directly visible and matched.[1]
The first scroll read whole survives as an 8 cm core because earlier attempts to open it failed. Technology read what hands had damaged.[1]
Primary sources are the 2026 preprint by Angelotti and colleagues, read in full; the Vesuvius Challenge’s own announcements of the 2023 Grand Prize and the June 2026 result; and reference sources for the library’s history. The 2026 result is a preprint not yet peer-reviewed. Source 6 is an interactive companion visual made for this case.
This case follows a library that could not be read without being destroyed — until a method, an open contest and specialist review together removed the trade-off. Many organisations hold the same kind of locked knowledge in legacy systems and archives. If you want to find a way to read yours without breaking it, write.