In AD 79 the eruption of Vesuvius buried the town of Herculaneum under pyroclastic surges hot enough to carbonise organic material in place. Among the buildings entombed was a seaside villa, now called the Villa dei Papiri, containing the only intact library known to survive from the classical world: roughly 1,800 papyrus scrolls, turned to fragile cylinders of carbon. Attempts to open them physically began in the 1750s with Antonio Piaggio’s mechanical unrolling frame and continued through chemical treatments and outright slicing; some scrolls were read that way, many were destroyed (Vesuvius Challenge, 2026). The remainder sat unreadable for two and a half centuries, a library you could hold but not open.

The idea that a sealed scroll might be read computationally is older than the current effort. Seales and colleagues demonstrated it on the En-Gedi scroll, a carbonised Hebrew text virtually unwrapped from a micro-CT scan (Seales et al., 2016), and laid the groundwork for Herculaneum specifically with the EduceLab-Scrolls dataset (Parsons et al., 2023). The Vesuvius Challenge, launched in 2023, turned the problem into an open competition with public data and prizes. In 2026 it produced the result the whole effort had been pointed at: a complete Herculaneum scroll, PHerc. 1667, virtually unwrapped and read without being opened (Angelotti et al., 2026a). The question is no longer whether a sealed scroll can be read. It is whether the process can be made reliable and automatic for every scroll, and that question is open.

Why it is hard

The pipeline has three stages, and each one is a genuinely difficult problem.

Scanning uses synchrotron X-ray micro-computed tomography, because carbon ink on carbonised papyrus offers almost no absorption contrast: ink and page are the same element. What signal exists lives in texture, morphology and phase, which is why the scans exploit phase contrast, the interference fringes that develop as a phase-shifted X-ray wavefront propagates to the detector. Even then, densely compressed regions of a scroll scatter the beam into a haze that no downstream algorithm may fully undo. Recent sub-micron scans traced this to the internal cavities of the carbonised fibres themselves (Vesuvius Challenge, 2026). Some of the damage, in other words, is done before any software runs.

Unwrapping is the stage that surprised me most. A scan does not give you sheets. It gives you a 3D grid of voxels, billions of them, with no information about which voxels belong to the same surface. Reconstructing the papyrus means predicting a surface probability for every voxel, usually with 3D U-Net variants (Isensee et al., 2021), then tracing an explicit mesh through that prediction, then flattening the mesh with low isometric distortion. A single local error can send a traced mesh jumping onto the neighbouring wrap with no easy way back, so the current system is semi-automated: humans inspect and correct at almost every stage. The labels those models learn from are themselves approximate, human-traced meshes that wiggle and drift, and the project now names label quality as one of its main bottlenecks.

Ink recovery is the strangest stage. Models are trained on detached fragments, where visible ink photographed in infrared provides labels for the corresponding CT data, then applied to sealed scrolls where no ground truth exists, bootstrapped with conservative pseudo-labels. Under favourable conditions the ink can now be segmented directly in 3D (Angelotti et al., 2026a), and separate work shows that surface topography alone, measured by optical profilometry, carries a usable ink signal (Angelotti et al., 2026b). But when a model finds no ink, nobody can yet say with confidence whether the signal is absent or merely unexploited. Telling “no ink” apart from “no ink recovered yet” is itself an open problem.

Why I think I can contribute

Three things changed this from something I read about to something I am doing.

The first is that the project went cloud-native. Full scroll volumes run to dozens of terabytes, but the open data is stored as chunked OME-Zarr (Moore et al., 2021) in a public S3 bucket, so tools stream only the region they need. You do not need a cluster to work on real scroll data. You need a good workstation and an understanding of the failure modes.

The second is that the organisers published a map of exactly where help is wanted, six concrete asks spanning the pipeline. Two of them fit my background almost suspiciously well. Mesh topology repair, which the project calls one of its highest-leverage contribution points, wants geometry processing, optimisation and C++: catching the holes, mergers and sheet switches that corrupt traced surfaces. Conservative fibre tracing wants classical computer vision: papyrus fibres physically define the row and column axes of each sheet, so a tracer that follows fewer fibres correctly is worth more than one that follows many with errors. Both are algorithm problems more than compute problems.

The third is a precedent buried in the project’s own write-up: an autonomous agent swarm, adapted from the open-source autoresearch project, found a training configuration that nearly doubled a cross-scroll validation score (Vesuvius Challenge, 2026). Agents have already produced a cited result here. My working method for several years has been exactly this, Claude as a tireless collaborator on codebases I direct, and this is the first research project I have seen that explicitly rewards it.

The threads it pulls together

I spent the early 2000s at Weta Digital, where the daily material of the job was geometry: meshes, surfaces, pipelines that turned messy scanned data into something an artist could work with, under deadlines set by cinema release dates. Before that I studied physics, which is why phase contrast imaging reads to me like an old friend rather than an obstacle. In the decades since I have written software more or less continuously, and for the last few years built with and around machine learning daily. On the side, I study Middle Egyptian hieroglyphs through the Egypt Exploration Society, because deciphering ancient writing is one of the purest puzzles there is, and I have a television series set in Pompeii in development, which makes the eruption of Vesuvius something I think about professionally as well as historically.

I do not believe in destiny, but I recognise convergence when I see it. Film taught me production pipelines and geometry. Physics taught me the imaging. Software taught me to ship. The hieroglyphs taught me why any of it matters: because a text that no living person has read is a voice waiting to be let back into the room. Most of the Herculaneum library is believed to be Greek philosophy, much of it by Philodemus, but nobody knows what else is in there. Every scroll is a sealed envelope from the first century.

What I am actually going to do

A new Apple Silicon workstation with 128GB of unified memory arrives shortly, which is comfortable for streaming inference and geometry work even where a consumer GPU’s memory would not be. The plan, deliberately modest: build the project’s tooling locally, stream a scroll region, reproduce a known topology failure case, and then commit to one lane, most likely mesh topology repair, with fibre tracing as the second candidate. I will write up what I find either way, including the dead ends. In a project this open, a well-documented failure is a contribution too.

Limitations, honestly

This post is a statement of intent, not of results. I have contributed nothing yet. Serious model training is out of scope for a laptop, however large its memory, and would mean rented cloud GPUs; the lanes I have chosen are chosen partly because they are algorithm-bound rather than compute-bound. The field around the deep learning stages is crowded with funded teams, which is a further argument for the geometry and tooling gaps. And there is a real possibility that my contribution ends up being tooling, documentation or labels rather than a named breakthrough. I can live with that. The scrolls have waited nearly two thousand years; they are owed patience and honesty about who is doing what.

References

Angelotti, G., Parsons, S., Nicolardi, F., Nader, Y., Johnson, S., Josey, D., Henderson, P., Schilling, H., Rudolph, J., McDonald, F., Dal Prá, E.R., Tafforeau, P., Mirone, A., Parker, C.S., Posma, J.P., Kyles, B., Vergara, C., Lavorante, A., Villa, R., Robustelli, M.C., D’Angelo, M., Del Mastro, G., McOsker, M., Fleischer, K., Chapman, C., Friedman, N. and Seales, W.B. (2026a) ‘Complete virtual unwrapping and reading of a rolled Herculaneum papyrus’, arXiv preprint arXiv:2606.29085.

Angelotti, G., Nicolardi, F., Henderson, P. and Seales, W.B. (2026b) ‘Ink detection from surface topography of the Herculaneum papyri’, Scientific Reports.

Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J. and Maier-Hein, K.H. (2021) ‘nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation’, Nature Methods, 18, pp. 203-211.

Moore, J., Allan, C., Besson, S. et al. (2021) ‘OME-NGFF: a next-generation file format for expanding bioimaging data-access strategies’, Nature Methods, 18, pp. 1496-1498.

Parsons, S., Parker, C.S., Chapman, C., Hayashida, M. and Seales, W.B. (2023) ‘EduceLab-Scrolls: verifiable recovery of text from Herculaneum papyri using X-ray CT’, arXiv preprint arXiv:2304.02084.

Seales, W.B., Parker, C.S., Segal, M., Tov, E., Shor, P. and Porath, Y. (2016) ‘From damage to discovery via virtual unwrapping: reading the scroll from En-Gedi’, Science Advances, 2(9), e1601247.

Vesuvius Challenge (2026) ‘Open problems: why reading every Herculaneum scroll is still a challenge’. Available at: https://scrollprize.org/2026_open_problems (Accessed: 20 August 2026).

Image: Herculaneum Papyrus 1521, British Library. Public domain.