A software engineer with a home AI lab pointed a chain of models at 4.35 million pages of Dutch East India Company archives and found a forgotten meteorite report, three lost rhinos, and volcanic eruptions nobody had recorded. He is not a historian. He built the pipeline over a weekend, inspired by historian Benjamin Breen, who used Opus 5.5 to find a 1615 eyewitness account of the dodo sitting unread for four hundred years. The documents were always there. Nobody had the time to read them.

Your plant has the same problem on a smaller timescale. Twenty years of quality reports, audit findings, deviation logs, and shift handover notes that nobody reads twice. Here is what AI archive mining actually looks like when you point it at your own records, and what it tends to surface first.

Every Plant Has a 400-Year Archive. Nobody Reads It.

Deviation reports, CAPA records, maintenance logs, shift handovers. Most of it gets read once, by the person who has to sign it, in the week it was written. After that it becomes storage. Your QMS holds a decade of evidence about why your lines fail, and nobody has opened most of it since the day it was filed.

Storage got cheap. Reading did not. A terabyte costs less than a shift supervisor’s lunch, while interpreting a single batch record still takes a trained human twenty minutes.

That gap is exactly what the GLOBALISE project ran into: millions of handwritten Dutch East India Company pages, turned into searchable text, still unread. Searchable is not the same as read. It took an AI pipeline chewing through 5.7 million passages to close the difference.

What Actually Came Out of 4.35 Million Pages

The dodo sighting that had been public for 400 years

Breen’s method was not exotic. He pointed Opus 5.5 at the GLOBALISE transcriptions of Dutch East India Company records and asked questions a historian would ask. What came back was a 1615 ship’s journal from Mauritius, where the crew “caught many tortoises, dodos”. A fresh eyewitness account of an extinct species, in a document anyone could have opened.

Waites read that post and his reaction was not admiration, it was recognition. He is a software engineer, not a historian, and he saw a workflow rather than a discipline. So he extended it: instead of chasing one question, he ran many open historical questions in parallel and chained models together to handle the filtering nobody wants to do by hand. One pass over 5.7 million passages, narrowed down automatically, with every surviving candidate checked against the original scan.

Why ‘the data already exists’ is the whole point

Not a single new record was created for this project. No expedition, no new digitisation, no fieldwork. The meteorite report, the rhinos, the unrecorded eruptions all came out of material that had been public and searchable before the work started. The only thing added was attention, applied at a speed humans cannot match.

That is the part worth sitting with. The bottleneck in historical archive research was never access. It was that reading four million pages carefully is a career, not a project, and nobody was going to spend a career checking whether a line about tortoises mattered.

Your plant is in the same position with better odds. The recurring failure mode on line three, the supplier whose deliveries correlate with your scrap spikes, the deviation that keeps reappearing under three different descriptions: that evidence is already sitting in your systems, written down by people who did their job properly. Nobody has read it together. AI archive mining is mostly the act of finally doing that.

The Six-Step Pipeline: Archive, Search, Filter, Read, Verify, Discovery

The architecture is the part you can copy. Six stages: archive, search, filter, read, verify, discovery. The 4.35 million pages get split into 5.7 million searchable passages, a cheap model called Jev culls the obvious noise, Claude Haiku reads what survives, and a stronger model checks every surviving candidate against original scans and catalogues before a human specialist ever sees it.

Nothing here is exotic. What makes it work is the ordering and the discipline behind two design choices that most internal AI projects skip entirely.

Cheap models triage, expensive models adjudicate

You cannot afford to send five million passages to a frontier model, and you do not need to. Most of what sits in an archive is irrelevant to any given question. The filter stage exists to throw away the 99% that will never matter, using a model that costs almost nothing per call.

The same economics apply to your deviation records. Run a cheap classifier across everything, let a mid-tier model read the shortlist, and reserve your most capable model for the handful of candidates that have already cleared two gates. Teams that skip tiering either burn the budget in week one or quietly shrink the corpus until the project stops being useful.

Every claim ends at a document you can open

The pipeline was built around a blunt premise: AI models can be confidently wrong. So every claim had to terminate in a real page, a scan of the actual handwritten letter or the actual printed newspaper, with an archive reference anyone can look up and read for themselves.

Apply that rule hard. Any finding your system surfaces should come with the batch record number, the shift note timestamp, the audit log line. No reference, no finding. If a quality engineer cannot click through to the source in under ten seconds, the output is a suggestion, and nobody acts on suggestions during a line stoppage.

Running the Same Play on Your Own Records

Start where the project started: with a ranked list of questions, not with a tool. The deep research step produced thirteen candidate questions scored on how complete the data was, how much AI genuinely helped, whether someone had already done it, and whether an answer could be checked against an original page. Run that same scoring on your own operation before you write a line of code.

Candidate questions that tend to score well in a plant:

• Recurring failure modes: the same root cause described four different ways across six years of non-conformance reports.
• Supplier defects that died quietly: flagged on receipt, noted in a shift log, never escalated to a supplier scorecard.
• Maintenance notes that called it early: free-text comments logged weeks before an unplanned stoppage.
• Audit findings repeating across sites: identical issues wearing different vocabulary in Mexico, Poland, and Ohio.

Choosing questions where the answer is checkable

The rule that made the archive work holds in quality: every claim has to end at a real document. Not a summary, not a confidence score. A deviation ID, a batch record, a dated line in a CMMS ticket that a human can open and read.

If a question cannot terminate in a source record, drop it. “Is our culture of quality improving” is unanswerable. “Which three defect codes appear in more than 40 percent of our 2021 to 2025 customer complaints, and in which specific records” is answerable, auditable, and defensible in front of a notified body.

What the ROI actually looks like: hours recovered, repeat issues caught

The economics sit in two layers. Filtering is where cost control lives, because a cheap model cutting 95 percent of the corpus means your expensive model reads a manageable remainder. Verification is where trust lives. Skip either and you get a pile of plausible nonsense no auditor will accept.

Measure two things. Investigation hours recovered per CAPA, and repeat findings caught before the next audit cycle instead of after. Both are countable from records you already keep.

The Advantage Shifts to Whoever Reads Their Own History First

The institutions that owned those archives had the data for decades. Universities, national libraries, research groups with funding and staff. They were beaten by one engineer with a home lab and a weekend, and that gap is the whole story. Reading capacity used to be the constraint. In 2026 it stopped being one.

What replaced it is harder to buy. The scarce skill now is framing a question precise enough to answer and building a pipeline that refuses to return anything without a document behind it. He put it plainly: AI models can be confidently wrong, so every claim had to end at a real document. That discipline is what separates a finding from a hallucination, and it is the part most internal AI efforts skip.

For a quality or operations leader, the move this quarter is small and specific. Three decisions, none of which require a budget cycle.

• One archive: pick the corpus you already trust and already own. Five years of deviation reports from one site beats a federated data lake you will spend two quarters negotiating access to.
• One question: something a line manager would actually act on if you answered it. Not “what insights are in our data.” That question has no verification standard, which means it has no answer.
• One verification standard: every output points to a document ID, a date, and a page. If a reviewer cannot open the source in under a minute, the result does not count.

Run that, and you learn more about your own failure patterns in six weeks than in the previous six years of filing them. Your competitors are sitting on the same unread history you are. The difference is going to be who opens it first and who keeps treating it as storage.

Archive mining with AI is not a research capability. It is an operational one, and the window where it counts as an advantage is open now.

Source: jessewaites.com