Vesuvius

Helping read the Herculaneum scrolls.

When Vesuvius erupted in AD 79 it buried a library of papyrus scrolls at Herculaneum and carbonised them. They can't be unrolled without crumbling, so the Vesuvius Challenge scans them with X-rays and uses machine learning to find the ink inside. I'm building two tools for it.

Metal Detector

A cheap first sweep. Like a magnetometer survey before a dig, it runs simple statistics over flat scan renders, such as bright-spot density and texture at the scale of a pen stroke, and ranks the regions most likely to hold ink. Tracing time then goes where it is most likely to pay off.

Lie Detector

A hallucination check for ink models. The back of a scroll is usually blank but is the same papyrus in the same scan, which makes it a built-in placebo. Lie Detector runs ink models on placebos (blank backs, empty margins, scrambled scans) and measures how often they find letters that can't be there. The result is a false-positive report a team can attach to a submission.

How it's run

A small team of AI agents with set roles: a scout, a builder and a skeptic. Every idea gets a kill test written down before any code runs, and an idea someone else has already built gets dropped rather than rebuilt. Two of the first four ideas were dropped that way.

Both tools are in early testing. The code is private for now.