Scale changed retrieval, not the burden of proof.
Jesse Waites describes an AI-assisted search across VOC records, newspapers, and ship logs. His pipeline turned millions of pages into passages, used a small decision model to filter candidates, sent a smaller set to a larger model, and then returned to original page images and specialist catalogues.[1]
That sequence matters. A semantic match is a retrieval result. A translated passage is an interpretation. A catalogue comparison tests whether the report is already known. A domain specialist can challenge the reading, context, place, and novelty. Do not compress these steps into one green "discovered" badge.
A model can reduce the pile. It cannot turn a candidate into a settled record by changing the label.
Build a known-item control before reading absence.
Waites tested the route against known events before he trusted a search that returned nothing. His controls included a known dodo passage and documented eruptions. The pipeline had to recover records already known to be present.[1]
This is the archive version of a proven-red test. Put known positives into the same corpus, transcription layer, embedding model, filters, and review path. If the system misses them, a null result says more about recall than history. Save the control set and its misses beside every absence claim.
The scan outranks the transcription.
GLOBALISE says it is using AI to transcribe more than five million handwritten VOC pages. It also publishes a searchable prototype and describes spelling variation as a search problem.[2] The transcription makes retrieval possible. It is not the archival object.
For the reported 1812 meteorite lead, Waites links the exact Delpher newspaper image. Delpher identifies the publication as the Java Government Gazette, dated December 19, 1812.[3] A review packet should preserve that resolver URL, publication, date, page region, transcription, translation, and the words that remain uncertain. That lets another reader inspect the mark instead of trusting a paraphrase.
Novelty is a search result with an expiry date.
The source article says the meteorite, rhino, and eruption findings remain candidates. Original pages and accessible catalogues were checked, but specialist review was still pending when the article was published.[1] That boundary belongs in the headline and the data model.
Use separate states for lead, page-checked candidate, catalogue-checked candidate, specialist-reviewed finding, and rejected lead. Record which catalogues were searched, when they were searched, and which names, dates, places, or spelling variants were used. A future catalogue update can then reopen the novelty check without erasing the earlier work.
The workspace is real. The historical verdict is not bundled with it.
Waites published Antiquity as a small Go repository that scaffolds a case workspace for AI-assisted archival research. We inspected the repository at commit e64fda33d081af09cf9594c03d64f4f7c68bd180. Its README describes folder conventions, a two-model funnel, known-item calibration, original-image review, and external catalogue checks.[4]
We did not build or run the CLI. The repository had no release tags and no detected license in the GitHub API response we read. Treat it as inspectable source, not as a packaged or licensed dependency. More important, a generated case folder cannot validate a historical claim. It can keep the chain of evidence from dissolving into chat.