---
id: "lib-2026-10-08-i-pointed-ai-at-400-years-of-historical-archives"
type: "library"
title: "I Pointed AI at 400 Years of Historical Archives"
lang: "EN"
org: "Jesse Waites"
source_type: "Practitioner post"
date: "2026-10-08"
week: "2026-W41"
url: "https://nosov.ai/library/lib-2026-10-08-i-pointed-ai-at-400-years-of-historical-archives/"
original_url: "https://jessewaites.com/blog/post/i-pointed-ai-at-400-years-of-archives/"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Detailed practitioner account of an AI archive-search pipeline with control tests and archive citations, but findings are unreviewed candidates from a single engineer."
summary_basis: "full-text"
original_minutes: "16"
tags: ["ai technology", "enablement", "tiered llm pipelines", "knowledge extraction"]
---

# I Pointed AI at 400 Years of Historical Archives

Tier B (Useful, limited): Detailed practitioner account of an AI archive-search pipeline with control tests and archive citations, but findings are unreviewed candidates from a single engineer.

Summary based on the full text.

## In short

Software engineer Jesse Waites extended historian Benjamin Breen's 1 October 2026 dodo find by building a multi-model pipeline over the Dutch East India Company archive, Dutch and American newspapers and ship logbooks. It surfaced an 1812 meteorite fall in Maharashtra absent from six catalogues, three Javan rhinos shipped toward Sri Lanka in 1738-1740, and three eruptions missing from the Smithsonian's Global Volcanism Program. All are candidates awaiting specialist review.

## Key numbers

- 4.35 million: Dutch East India Company archive pages searched (GLOBALISE transcriptions, 1600s-1790s)
- 5.7 million: passages embedded in one twelve-hour overnight run on a home GPU
- about $3: cost for a small filter model to read 59,000 elephant mentions
- 769 to 3: eruption reports found, narrowed to three candidates missing from the Smithsonian list

## What it reports

- Reading the VOC pages alone at two minutes a page, full-time, would take about 70 years; embedding them took one overnight run.
- Pipeline: semantic embeddings, then a small "System One" filter model (Jev) at a few cents per million words, then Claude Haiku to translate and extract, then Claude Code to check each transcription against the original scan.
- Control rule: before searching for new events, the pipeline had to find known ones (Breen's dodo, Laki 1783, Tambora 1815 and a dozen others).
- Meteorite: an officer's letter of August 1812, reprinted in the Java Government Gazette of 19 December 1812, describes a stone of about four pounds; if confirmed it predates Maharashtra's earliest recorded fall by 26 years.
- Rhinos: one died in Colombo in 1738; two aboard the Loverendaal, sailing 5 July 1740, died on the voyage, with sworn crew statements.
- Eruptions: Gamkonora 3 February 1722, Ciremai early 1712, Slamet February 1780; the 1711 Awu eruption matched the Smithsonian record to the day as a check.
- The search for the unlocated 1808/09 eruption, including all 633 articles of Batavia's first newspaper, found nothing.

## How they got it

A deep research assistant first ranked 13 open questions by data availability and checkability. Waites and Claude Code built the pipeline on a home GPU, with Waites choosing leads and stopping points. Candidates were checked against original scans and specialist catalogues; curators and Kees Rookmaaker have been contacted.

## Limits

- None of the finds has yet been confirmed by catalogue maintainers.
- Newspaper "oddities" were not checked against original pages.
- Model guesses about volcano identity were often wrong, one by 1,500 km, and needed manual placement.

## Why it matters for companies

The pipeline is a template for searching any large, messy document estate: contracts, maintenance logs, old correspondence. The cost structure is the lesson: a cheap filter model screens tens of thousands of hits, a larger model reads only the survivors, and a person decides. The control test, proving the search finds known items before trusting empty results, applies directly to enterprise search and audit work.

Original: https://jessewaites.com/blog/post/i-pointed-ai-at-400-years-of-archives/