I Pointed AI at 400 Years of Archives. It Found a Forgotten Meteorites and More

A software engineer built an AI research pipeline to trawl millions of digitized early-modern records and surface neglected eyewitness reports that can be verified against original scans. Starting from a validated prompt list of historical questions, the system turned GLOBALISE’s 4.35 million VOC pages plus centuries of digitized newspapers and ship logs into semantic fingerprints, letting the pipeline search meaning rather than brittle keyword variants. A cheap, narrow “System One” classifier (Jev) filtered millions of candidate passages into a few dozen leads; larger models translated, extracted dates and places, and an agent opened original manuscript or print scans to confirm transcriptions. Controls required the pipeline to rediscover known events (for example well-known eruptions and the dodo sighting) before new claims were accepted, and the operator remained actively involved at every stage.

Applying this workflow produced concrete discoveries: an 1812 Bombay Gazette report reprinted in the Java Government Gazette describing an iron-rich stone that fell near Pandharpur - boom, a shallow crater and a black fusion crust - an otherwise unlisted meteorite that would become the earliest recorded fall in Maharashtra. The search also recovered three Javan rhinoceroses intended as diplomatic gifts to the King of Kandy that died en route (one in Colombo stables in 1738, two aboard the Loverendaal in 1740). Alongside those finds, the pipeline surfaced previously overlooked newspaper items and reports pointing to unrecorded volcanic eruptions, expanding the primary-evidence base for several niche historical questions.