An engineer had AI agents comb old archives and newspapers, and they turned up an 1812 meteorite fall that none of six meteorite catalogs record

Waites read Breen's dodo post and thought, "I can do this." The first step was asking a Deep Research assistant which historical questions could likely be answered with material already online. It came back with 13 candidates. Next, Waites built a pipeline with Claude Code: index with embeddings, search by meaning, filter with a small judge model called Jev, have Claude Haiku translate and pull out dates and places, then have a Claude Code agent check the results against the original scans and match them against scholarly catalogs of meteorites and volcanoes. jessewaites

Waites indexed 4.35 million pages of Dutch East India Company records, or 5.7 million passages, on a single GPU running overnight for 12 hours. By Waites's math, a person reading 8 hours a day, 5 days a week, at 2 minutes a page would need about 70 years to get through it. Filtering was cheap too. A search for animal records pulled up 59,000 passages mentioning elephants, and having Jev read and filter them cost about $3. For meteorites, the searches used sentences like 'a heavy stone fell from the sky with a loud report,' according to JTPCK's case study page. jessewaites +1

Before the real search, Waites had the system find events that were already known. Checking whether Breen's dodo record, the 1783 Laki eruption, and the 1815 Tambora eruption came up confirmed that the search was working. jessewaites

There were three results. The first is the 1812 meteorite fall on the page at the top. It came from old newspapers, not the Company archives. Waites searched old newspaper articles alongside the VOC records, and the meteorite story ran in one of them, the Java Government Gazette. The original page can be viewed on Delpher, the National Library of the Netherlands' historical newspaper service. The fall is not in any of the six meteorite catalogs. If confirmed, it predates the earliest known meteorite fall in Maharashtra, India, where Cokurrgaum is located, by 26 years. jessewaites +1

The second is three Javan rhinos sent as gifts to the king of Kandy in Sri Lanka in 1738–1740. One was shipped in 1738 and two in 1740 aboard the Loverendaal, and all of them died before arriving. These shipments did not appear in any of the rhino record lists Waites checked. The third is volcanic eruptions. Filtering 769 eruption reports left three eruptions that are missing from the Smithsonian volcano catalog: Gamkonora in February 1722, Ciremai in early 1712, and Slamet in February 1780. The original rhino and volcano records are VOC documents held by the National Archives of the Netherlands. jessewaites

All three are still candidates that experts have not reviewed. Waites has contacted a meteorite curator, rhino researcher Kees Rookmaaker, and volcanologists, and is waiting to hear back. Waites also tried to pin down the great eruption of 1808–09, whose source volcano is still unknown, but found nothing relevant even after the agents read all 633 articles in the first newspaper published in Batavia (today's Jakarta). jessewaites

While preparing the dodo chapter of a book on animal extinction, Breen had an Opus 5.5 agent search the VOC records published by GLOBALISE. GLOBALISE is a project by Dutch research institutions that transcribes handwritten Dutch East India Company documents with AI and makes them available for free. Breen used GPT-6 as well. As Breen describes it, agents like Opus 5.5 spin up copies of themselves to split the reading of sources in several languages, and given an API key, they run embedding searches on their own. On this project, the agents multiplied to around a dozen. globalise.huygens.knaw.nl +1

The passage the agents found is in the April 1615 logbook of the ship Wapen van Amsterdam. The crew wrote that on Mauritius they "caught many tortoises, dodos [dodeersen], and some geese and parrots." The original is in the VOC archives at the National Archives of the Netherlands. The timeline of dodo sightings had no records for 1611–16. Breen sees this passage as a new record for that gap and wrote, "I do think that this is a publishable result." In the comments, Breen added that to the best of the author's knowledge, the April 1615 passage had never been reported before. resobscura.substack +1

The agents also reread known records in the originals. In a 1638 document, the word 'velthoenderen' (probably the red rail) had been rendered as 'perdrix' (partridge) in an 1890 translation by a French scholar. Opus 5.5 went back to the original manuscript and corrected that translation. Neither the 1615 passage nor this correction has yet been formally reviewed by specialists in the field. resobscura.substack

This case is not about searching a pile of documents. It is about reading one document nobody could read. Church gave GPT-6 Astra a single scan of a cipher letter addressed to General Marmont, published in a French military journal in 1969, and ran it for about six hours. Working from that one scan, the agent did everything from transcribing the symbols to breaking the cipher. Church wrote, "I spent six hours of model time and a few evenings of my own." carter.church

The letter consisted of 1,300 cipher units drawn from 155 distinct symbols. Only 33 symbol meanings had been published before, and they covered 435 of the 1,300 units. The agent worked out the remaining two-thirds or so from the letter itself. It read the full letter in French and English, and corrected the date, misprinted in the 1969 reprint, to late March 1809. carter.church

Church also published a solution package with the key (key.tsv), the transcribed ciphertext, and verification scripts (verify.py, reproduce.py), so anyone can rerun and check it. Israeli outlet Calcalist reported that Satoshi Tomokiyo, who runs the cipher research site Cryptiana, has marked the cipher as solved. carter.church +1