The lede
Using Opus 5.5 to discover a new eyewitness record of the dodo
Benjamin Breen, Res Obscura · made the Sunday
The method is the thing, and it is unglamorous enough to copy: take a real research question you already have standing, find a large well-edited corpus (here the GLOBALISE archive of Dutch East India Company records), embed it, use semantic search to surface candidates, and let agents read the hits in several languages and rank them for a human. What came back was a 1615 ship's log from the Wapen van Amsterdam noting a crew that 'caught many tortoises, dodos, and some geese and parrots' — filling a documented gap in the dodo timeline — plus a 1638 reference to the extinct red rail that a French scholar mistranslated as 'partridges' in 1890 and nobody re-checked. Breen is just as clear about the failure mode: left alone, a dozen agents drill into Inca khipu and produce findings he has no way to evaluate, because models are bad at judging significance and cannot ask the right question in the first place. The conclusion is the one worth carrying into your own work — the bottleneck moves off the finding and onto the attention of the few people qualified to tell you whether it matters.