Project summary

The Vesuvius Challenge (scrollprize.org) pays $50,000 per scroll for the first ten legible letters found inside a carbonized Herculaneum scroll, up to ten scrolls and $500,000 in total, with a deadline of June 25, 2027. In September 2026 the eligible set was widened from the thirteen 2027 Grand Prize volumes to twenty-three, adding ten scrolls in which no text has ever been read. No ink has been found in any of them yet.

The ink models are public and they work. The obstacle is everything around them: turning a scan into an ink map still needs Linux, a C++ toolchain, terabytes of local storage and a data-center GPU, so very few people can run the search at all, and the eligible scrolls are being searched slowly by a handful of teams.

I built a toolkit that does it on one RTX 3090 under Windows, from the organizers' published data alone: https://github.com/gmDevi/vc-windows-tools, MIT-licensed. It has reached all eighteen eligible scrolls that publish tracks: seven swept deep (69-165 windings each) and eleven given a first thin-band screen, 1,930 scored renders over 965 windings in total. I am asking for $6,000, almost all of it GPU time, to finish the pass over the whole eligible set and publish every map, positive or negative.

What are this project's goals? How will you achieve them?

The goal is one complete, public, reproducible ink search over every eligible scroll, instead of the partial sampling that one consumer card allows. Four steps:

1. Fit spiral meshes for every track-ready eligible scroll over its full height. A 30,000-step fit over one z-band measures 1.0-2.4 GPU-hours on a 24 GB card (seven fits, mean 1.5); a full scroll needs six to fourteen such bands plus two or three short restarts to settle winding count and direction, so roughly 10 GPU-hours a scroll and about 250 across the set.

2. Render and sweep every winding of a fitted mesh with all four published ink models, both layer directions, over the whole circumference, in a 7.5 mm z-band. The per-winding cost is measured: 120 timed windings, mean 5.2 minutes each (median 4.6, range 1.3-12.8) for one model in both directions. That is about 31 GPU-hours per scroll-band, so roughly 1,100 GPU-hours buys one band on all twenty-two track-publishing scrolls plus a second band on the ten that look most promising.

Being explicit about what that is not: these scrolls are 116-126 mm tall, about sixteen such bands each, so a full-height pass over the whole eligible set at this density would be nearer 11,000 GPU-hours. This grant funds about a tenth of that, and I will publish exactly which bands were swept rather than implying full coverage. The 5.2-minute unit is also measured only on the flat 9 um model; the other three are validated on the Paris 4 control but not yet timed here, so the four-model multiplier is an assumption I will replace with real numbers in the first published sweep.

3. Fine-tune the 2.4 um ink model on PHerc1203 and PHerc0846A, the two eligible scrolls scanned at the model's native resolution (2.403 um), and re-sweep both. Each 2.403 um volume covers 15,137 slices, about 36 mm of height rather than a whole scroll, which bounds the work: about 100 GPU-hours, budgeted rather than benchmarked, since 2.4 um inference costs far more per unit area than the 9 um sweeps measured so far.

4. Publish all of it: per-winding score tables, previews, and the fitted meshes, so other contributors and the organizers can look at the same maps instead of taking my word for them.

The outcome is either the first ink in a new scroll, which is a $50,000 prize and a text unread for 1,946 years, or a complete public negative with maps, which tells everyone that the current models need new training data rather than more sweeps. Both are worth having. Only the first is paid for by the challenge, which is why the search itself goes unfunded.

The reason to do this now rather than wait is that the cost has collapsed. The tooling exists and the data is public, so about 1,500 marketplace GPU-hours (RTX 4090 / A100 at $0.4-2 per hour) now cover a first banded pass over a set for which no winding-by-winding ink sweep has been published outside the organizers' team, as far as I can find. Every fix flows back upstream under an open license, so the money widens the funnel for every contributor rather than paying for one more closed attempt.

How will this funding be used?

• $3,000 - about 1,500 GPU-hours on marketplace GPUs (RTX 4090 / A100, $0.4-2 per hour) for steps 1 to 3 above.

• $2,500 - three months of part-time maintainer time: running and babysitting the fits and sweeps, keeping the toolkit and the upstream recipes current, answering other contributors.

• $500 - object storage and egress, so the meshes, renders and score tables stay downloadable.

At the $1,500 minimum it is compute only: the PHerc1203 fine-tune and full sweep, plus full-height fits and sweeps of the two cleanest meshes so far, PHerc0125 and PHerc0211. The results get published either way.

Who is on your team? What's your track record on similar projects?

Just me: Gennaro, a software engineer who lived in Portici (confining neighbour to Ercolano, where the scrolls where first opened and destroyed, which ex post kickstarted the prize!) and has always been fascinated by what's in the scrolls.

vc-windows-tools answers "does model X see ink on sheet Y of scroll Z?" from the organizers' published data products alone, on one consumer GPU under Windows:

- it streams only the 128^3 CT chunks a sheet touches from the organizers' S3 buckets, with no local copy of a volume;

- four renderers produce face-on sheet stacks: height fields grown from the published surface predictions at 9 and 2.4 micron, axis-aligned boxes, tilted sheets, and windings of a fitted spiral mesh (a 28 mm x 14 mm sector needs about 360 chunks, 0.7 GB, three minutes);

- it drives the four published ink models (flat 9 um, canonical 2 um ResNet3D-152, DINO-guided 3D U-Net, dense DINO ink-likeness) without S3 access or containers;

- it sweeps a whole scroll or a whole fitted mesh unattended and scores every render against a calibrated noise ceiling (the Paris 4 control crop with clear letters gives a fraction of pixels above 200/255 of 0.0112; speckle noise reaches about 0.004);

- it makes the official spiral fitter run under WSL2 with a one-line recipe, and documents two configuration traps that silently produce empty meshes;

- it ships the per-scroll starting parameters the fitter needs and the published data does not include - winding counts, winding senses and estimated umbilici - for every eligible scroll that publishes tracks, including nine of the ten added to the First Letters set on September 10, 2026, derived the same day and published the next. The tenth, PHerc0846B, publishes no track data, so nobody can fit it yet.

It is validated against a known answer: the pipeline reproduces the organizers' ink map on the PHerc. Paris 4 control segment, crisp Greek letters forward and noise in the reverse direction.

Results so far: 1,930 scored renders over 965 fitted windings in 26 batches, across eighteen scrolls (PHerc0125, 0175A, 0175B, 0191, 0211, 0257, 0268, 0306B, 0343, 0358, 0483A, 0483B, 0490A, 0490B, 0800, 0813, 0826, 0846A), every render scored with the flat 9 um model in both layer directions, with per-render JSON scores and PNG previews logged. The other three models are wired up and validated on the Paris 4 control but have not been swept at scale; that is what step 2 buys. All negative at the calibrated threshold except four renders in three early batches, each checked by eye and written up in the index: three came from fits that later runs superseded, and one is a crushed multi-layer region with dense speckle and no letter forms. As far as I can find, that is the largest non-detection sweep published from one consumer GPU. It does not show there is no ink: it shows that these models, at this resolution, on these bands, do not find any, which is consistent with the organizers' own note that ink remains elusive in the new data.

Four pull requests carry the fixes and recipes back to the organizers' repository, open as of September 11, 2026, none reviewed by a maintainer yet and #1735 parked as a draft under the repository's three-open-PR cap for outside contributors: https://github.com/ScrollPrize/villa/pull/1732, /1735, /1736, /1745. The work is entered in the September 2026 progress prize.

The clearest outside signal is that someone else uses it. An independent developer, GitHub user evilaliv3, wrote publicly on one of those PRs that he is already using the toolkit, and has since sent three measured patches to its umbilicus estimator, each with its own before-and-after measurement. Two of the three reproduce exactly when re-derived from the published data and are merged with attribution. Reviewing his work also turned up two errors of my own in the README, which I measured properly and corrected.

The full report, with the validation table, the sweep scoreboards and the negatives, is at https://github.com/gmDevi/vc-windows-tools/blob/master/report/REPORT.md

What are the most likely causes and outcomes if this project fails?

• Every sweep stays negative. This is the likely outcome for at least some scrolls. It gets published as a scored, reproducible negative with maps, which is still the most complete public test of the current models on the new scans.

• Spiral fits fail on the crushed or folded scrolls. PHerc0826 and PHerc0358 already show dense multi-layer regions. Mitigation: fit in z-bands, as I do now, and report per-band coverage rather than claiming a whole scroll.

• Marketplace GPU prices spike. The plan is sized at the top of the price range and the sweeps are resumable, so it degrades to fewer scrolls rather than failing.

The failure I would most want to avoid is a false positive: announcing ink that turns out to be noise. That is what the calibrated threshold, the reverse direction control and the published previews are for: every claim can be checked against the same data by anybody.

How much money have you raised in the last 12 months, and from where?

None. No grant has been awarded and nobody has given me money for any of this. Everything so far ran on one RTX 3090 I already owned plus free tiers (Kaggle, Colab, Paperspace).

I have applied widely, though, and a bare "nothing" would be misleading, so here is what is pending as of September 11, 2026: Emergent Ventures ($12,000, a six-month version of this same plan); the Vesuvius Challenge September progress prize; Google TPU Research Cloud; Lambda Research; Anthropic AI for Science; the OpenAI Researcher Access Program and the OpenAI Codex Open Source Fund; the Vercel Open Source Program; a Solana Foundation grant for an unrelated project of mine; and small credit programs at AMD and Saturn Cloud. Several more are drafted but not yet sent. Most are compute credits rather than cash, and most will be declined.

If any of them lands I will say so on this page and narrow this project to whatever is still unfunded. I will not bill the same GPU-hours twice.