Using Opus 5.5 to discover a new eyewitness record of the dodo

Using Opus 5.5 to run semantic searches over a large, well-edited corpus of Dutch East India Company records produced what appears to be a previously unnoticed eyewitness entry for dodos dated April 1615. The search pipeline used embeddings to surface candidate passages, then had frontier models read and rank matches for human review. One manuscript - a ship’s log probably by Isbrant Cornelisz van Petten of the Wapen van Amsterdam - records that the crew caught “many tortoises, dodos (dodeersen), and some geese and parrots” on Mauritius. That entry fills a gap in the dodo timeline (1611-1616) and sits alongside another recovery: a corrected reading of a 1638 reference to red rails (velthoenderen) that had earlier been mistranslated as partridges.

The broader claim is that contemporary research agents can produce novel archival finds by exhaustively trawling massive datasets, but they exhibit an epistemological oddity: they are superb at blind, relentless retrieval yet poor at framing questions, judging significance, and reliably tracing chains of transmission. Examples include a promising, unresolved hypothesis linking a 1616 Portuguese Jesuit account (called an “ostrich”) to the famous Mughal court dodo painted by Ustad Mansur, and an entertaining archival vignette about an English captain stealing giant tortoises. Human specialists remain essential to validate, contextualize, and decide which AI-flagged leads are historically consequential.