CV Meets Archaeology: Charred Herculaneum Scroll Yields First Letters, $50,000 First Letters Prize Awarded

Forum topic · QianXun · 2026-11-04

Summary

The Vesuvius Challenge awarded its $50,000 First Letters Prize for scroll PHerc. 306B to Liam Warner, a freelance software engineer who trained on a single RTX 4090, and Nicola Bodill, a University of Sydney classics master's student who hand-labeled hundreds of hours of ink data. Their pipeline used ESRF micro-CT scans, VC3D's GrowPatch surface segmentation seeded by a public spiral-fitter, and ink detection combining TRACE v3B with a fine-tuned A-hn model (derived from Youssef Nader's v8-in). The team read the Greek phrase "to di hauto haireton" ("choiceworthy for its own sake"), resolving a long-debated mu/nu ambiguity, with parallels in Aristotle, Sextus Empiricus, and Philodemus' On Anger. The reading also established that PHerc. 306B is the upper half of a scroll whose lower half is catalogued separately as PHerc. 306A. Notably, 2026 saw two First Letters prizes: Erwin Nieuwlaar won for PHerc. 343 in October using open-source recipes from the earlier win. The $800,000 Grand Prize for reading 70% of a scroll is targeted for 2027, bottlenecked by scan speed and cost.

Key points

The prize: The Vesuvius Challenge's $50,000 First Letters Prize for PHerc. 306B went to two winners: Liam Warner (freelance software engineer, northern England) and Nicola Bodill (classics master's student, University of Sydney). Of 25 submissions on this scroll, eight came close, with several teams reaching 7–9 recognizable characters. The threshold: 10 recognizable characters within 4 cm².
The reading: Everyone got stuck at the same spot — after τοδιαυτο, faint strokes were read as either mu or nu for nearly a month. Warner and Bodill's imagery clarified that the second alpha's central stroke rises diagonally from the base of the left stroke, matching the first alpha, settling the reading as αι. The full phrase reads τὸ δι᾽ αὑτὸ αἱρετόν — "choiceworthy for its own sake" — with close parallels in Aristotle, Sextus Empiricus, and Philodemus' *On Anger*.
The technical pipeline:

ESRF synchrotron micro-CT scan

VC3D GrowPatch segmentation

Ink detection

TRACE v3B

A-hn

Clarifying the "one RTX 4090 took $50k" framing: Warner only tried ink detection because he had a spare consumer GPU; Bodill had hand-labeled hundreds of hours of ink data since 2023. The prize was shared. Also, the winners' actual claim box was 38.0 × 9.56 mm = 3.60 cm² (under the 4 cm² cap), scanned at 8.64 µm voxels, 116 keV.
A cataloguing correction: 306B is the upper half of the original scroll; 306A is the lower half — one scroll had been registered as two for most of a millennium, raising 306A's scan value.
Don't confuse two prizes: The winners page also lists an October 8, 2026 First Letters prize to Erwin Nieuwlaar for PHerc. 343 (a secondhand RTX 3090 plus ~$300 of rented GPU, averaging four models, reading πᾶν τὸ γινόμεν[ον]). The 306B repository explicitly credits Nieuwlaar's open-source recipes — one open-source win directly seeding the next.

Why it matters

The Vesuvius eruption of AD 79 carbonized the Villa of the Papyri's library at Herculaneum; ~1,800 scrolls excavated in the 1750s crumble at a touch. Machine vision read the first word from these scrolls in 2023; the first letters of PHerc. 306B followed in this award. A consumer GPU plus one person who reads Greek beat well-resourced institutional teams.

What's next

20 more scrolls qualify for First Letters; 13 already have submissions.
The $800,000 Grand Prize targets reading 70% of a scroll with minimal human intervention by 2027.
The bottleneck isn't models — it's scanning: reliable ink detection needs 8–9 µm resolution scans, and fast, cheap scanning at that quality doesn't exist yet.

Sources & caveats: Vesuvius Challenge official announcement (scrollprize.substack); winner names, prize structure, model composition, and the 306B/306A determination all follow official statements. The Greek reading awaits further scholarly review and is not yet settled. Greek transliterations are loose Latin renderings.