Provenance
Chain
1. Images. Nationaal Archief, The Hague, archive 1.04.02 (VOC),
"Overgekomen brieven en papieren", 1610–1796.
2. Transcription. GLOBALISE (Huygens Institute, KNAW) ran the Loghi
handwritten-text-recognition engine over 4,802,212 page images and published the
result as "VOC transcriptions v2", doi:10.34894/50ZYWS, CC0. Not human-checked.
3. Selection. Inventories were chosen by parsing the Nationaal Archief EAD for
1.04.02 into 15,064 inventory descriptions and intersecting those with the
GLOBALISE deposit, filtered by place name and date.
4. Translation. Google Gemini 3.7 Flash on Vertex AI, August 2026.
Segments of ~6,000 characters split on paragraph boundaries, temperature 0,
reasoning disabled, one instruction: translate, repair obvious transcription
errors silently, add no commentary, preserve names, dates, numbers and units.
5. Anchoring. Page identifiers were recovered after translation by walking the
raw transcription and its segment boundaries in step, then verified against
known passages.
Known weaknesses
Two lossy steps, not one. The Dutch is machine-read from handwriting and the
English is machine-translated from that. Errors compound. Numerals survive better
than words; unfamiliar proper nouns survive worst.
Names are unstable. The Dutch clerks rendered South Asian names by ear, and the
translation preserves what it was given. Balwant Singh appears as "Bolwantsing",
Ghazipur as "Paziepoer" and "Gaziepoer", Chunar as "Chunagoor". Searching for a
modern spelling will miss them. Search the Dutch as well as the English.
Gaps are not random. 189 segments of roughly 102,000 initially failed because
Vertex's safety filter blocked the prompt outright. Colonial records describe
violence, coercion, punishment and slavery, so the filter fires hardest on the
passages a historian most wants. Those segments were recovered by narrowing the
window or routing to a second model, and anything still missing is marked inline as
[UNTRANSLATED SEGMENT] and listed in gaps.csv. **The absence of a subject in
this corpus is not evidence of its absence in the archive.** Check gaps.csv
before drawing a negative conclusion.
A worked example of a false negative. The Raja of Benares, Chait Singh, does not
appear anywhere in this corpus under any spelling tested. His father does, as
"Bolwantsing", in Dutch copies of the 1764 farman and the Treaty of Allahabad
(inventory 3509, scans 0921–0922). One absence is real; the other was a spelling
problem. Test your instrument on something you know is present before you report a
silence.