The Herculaneum scrolls: how virtual unwrapping reads a scroll that cannot be opened

The Herculaneum scrolls were baked solid by Vesuvius in 79 AD. They are cylinders of charcoal. Unrolling one destroys it, and the destruction was proved repeatedly across two centuries of trying.

Virtual unwrapping is three separate problems that get discussed as one, and only the third is really hard.

Stage 1: scan

X-ray computed tomography. Rotate the scroll, take thousands of cross-sectional slices, stack them into a 3D volume. Standard technology, applied at high resolution to an unusually fragile object.

Stage 2: segment and flatten

Inside that volume, the papyrus is a single continuous sheet wound into hundreds of tightly packed layers, squashed and torn. Segmentation traces one layer through the volume as a surface. Flattening then maps that crumpled surface onto a plane.

This is genuinely difficult geometry, particularly where layers touch or where the sheet has torn and rejoined. It is also, importantly, geometry: it can be checked by whether the recovered surface is continuous and physically plausible.

Stage 3: find the ink, which is the actual problem

The ink is carbon black. The papyrus is carbonised, which is to say also carbon. To a scanner looking for density contrast there is essentially nothing to see, and to a human eye looking at the flattened surface there is nothing at all.

What survives is a very slight difference: a thin dense layer sitting on the surface where a pen passed, altering how X-rays travel through those specific points by a tiny margin. A neural network trained on examples where the answer is known learns to flag it.

And that is exactly where the method becomes a statistical claim rather than a measurement, because a model trained to find faint patterns in noise will find faint patterns in noise.

Why keeping the stages separate is the safeguard

If geometry and ink detection ran as one opaque step, a model could in principle bend the recovered surface towards wherever it thought letters ought to be. Keeping segmentation independent of ink detection means the surface is fixed by physical continuity before anything starts looking for text on it.

There is a second check, and it is the strongest one available. Noise does not obey Greek grammar. Recovering passages that read as coherent Greek, in a hand consistent across the sheet, saying things a papyrologist recognises as the kind of thing these texts say, is extremely hard to fake by accident.

The number that is still missing

Ask what the character-level error rate is and you will not find one.

That is not a criticism of that paper, which says plainly what it did. It is a description of the field. In 2025, two years after the Grand Prize, careful teams still report how many letters they recovered and not how often they were wrong.

It is the same shape as a lidar survey reporting 400 candidates and no recall figure. Both are counts of hits. Neither has a denominator. And in both cases the missing number is the one that would let you compare the result with anybody else's.

A licensing footnote that says something about the field. The En-Gedi paper that introduced virtual unwrapping is fully open access, gold status, with a real Creative Commons licence. The licence is CC BY-NC, which forbids commercial use, so it is the best figure in this whole story and it cannot be printed in a book that is sold.