A library stood beneath this mountain.
The Vesuvius Challenge
Beneath
Vesuvius.
I’m building tools to look for ink in scrolls too fragile to open.
Then Vesuvius erupted.
The scrolls turned to carbon.
Now X-rays can see inside.
My part: testing whether a mark is ink or just papyrus.
The challenge
Reading a scroll
without opening it.
Vesuvius buried Herculaneum and carbonized hundreds of papyrus scrolls. Unrolling them can destroy them.
The Vesuvius Challenge uses CT scans and machine learning to read them without opening them.
01 / 03
Scan the scroll.
CT imaging captures thousands of slices through the papyrus.
Flatten a sheet digitally.
Software traces one layer through the folds.
Look for ink.
Models flag candidate marks for closer inspection.
What went wrong
The models agreed.
The marks weren’t ink.
Several models highlighted the same spots. A closer look explained why.
They were picking up fibers, seams and torn edges.
The pivot
Start with
known writing.
I needed to check the method before using it on unread scrolls.
Use a scroll already read.
Experts have published its text, so misses are easy to spot.
Pin each character to the surface.
Each gets exact coordinates anyone can inspect.
Trace it back into the scan.
That ties the mark to real papyrus, not just an image.
Look for other explanations.
The plan uses chance tests, blank regions and alternate masks.
What the checks
found.
The checks covered known text, reproducibility and false leads.
Known characters traced to the scan.
PHerc0139 linked known characters to surface coordinates and their CT volume.
PHerc1667 rebuilt end to end.
Every hash checked. The output rebuilt from pinned source and binary.
Render geometry explained the apparent ink.
All five PHerc1447 leads failed the structural test.
Cohors Vesuviana
How I
split the work.
I split the work across six AI roles, with Romulus coordinating.
OperatorRomulusRoutes tasks, checks results and brings me a recommendation.
Minerva
StrategyPicks the next question.
Vulcan
EngineeringBuilds tools and logs runs.
Veritas
FalsificationTries to break each claim.
Mercury
IntelligenceTracks outside work.
Plutus
CostTracks compute and cost.
Cassandra
RiskFlags risks early.
Research snapshot · August 29, 2026
Where things stood
on August 29.
Five PHerc1447 leads
The renderer explained every gap. All five closed.
STRUCTURAL_EXPLAINEDPHerc800 surface gate
The metric was invalid, so it failed before reaching the unread scroll.
FAIL_METRIC_INVALIDPHerc1667 canonical chain
Every hash and coverage figure reproduced.
CANONICAL_CHAIN_PASSM1v2 offset scan
The next test was a possible 2.3-voxel offset between the surface labels and the sheet center.
Result pending at snapshot
OFFSET TESTThe question
The failed gate raised a question: were the surface labels aligned with the sheet center?
No offset result recorded in this snapshot.