The Archive That Learned to Answer

Segment one: the problem

On screen: the project folder, slowly scrolling the top level.

Most digital libraries are filing cabinets. You search, you get a hit, and you're on your own. I wanted something else. I wanted an archive that responds.

The subject is Gilles Deleuze, the French philosopher. He's hard to hold in your head. He wrote about thirty books, alone and with Félix Guattari. He also taught for sixteen years at Vincennes, and many of those seminars were recorded. The same concept comes back in a book from 1968, a seminar from 1981 and an interview from 1988. Each time it has shifted a little.

So the question I brought to Claude was simple. Could we build an archive that keeps all of that in view at once, where any concept has many paths to it and nothing is ever made up?

Segment two: the books, handled like evidence

On screen: the INVENTORY file, then the toolkit scripts folder.

We started with the published books in English. Every book went through the same pipeline: extract, convert, check, clean up, then two rounds of readability scans.

One rule governed all of it, and we called it zero content mutation. Whitespace can be tidied. Letters, words and punctuation are never added, removed or changed without my written approval, and every approved change goes into a log. Each finished text gets a fingerprint, a hash, recorded in the inventory. At the start of every session the archive checks itself against those fingerprints and reports any drift. It doesn't quietly fix anything.

Here's one example I'm proud of. The English translation of Deleuze's book on Foucault only existed as two imperfect scans, from two publishers. We rebuilt it by comparing them. One scan decided the characters. The other decided how the footnotes were numbered. That book closed the first phase of the project on May 30, 2026.

Today the English shelf holds twenty-eight books, about 2.76 million words. A French shelf of twenty-three books sits beside it. Those were digitized from printed copies I own, for personal research use.

Segment three: the seminars and the voice

On screen: the seminars folder, one seminar opened to show its English, French and audio subfolders.

The books are only half of it. The seminars are where Deleuze built his concepts live, under questions from his students. We filed 492 seminar transcripts across thirty-four seminars. We gave the seminars their own citation system, down to the week of a single lecture.

Then we went after the recordings. We drew on three public archives: Purdue University, La Voix de Gilles Deleuze at Paris 8, and Gallica at the French National Library. By May 11 the archive held 1,132 audio files, about sixty gigabytes, across eleven seminars.

Among them is Deleuze's last lecture, given on June 2, 1987. That's the close of a thirty-nine-year teaching career, and a researcher can now cite it the same way they'd cite a page of a book.

Segment four: when the AI was wrong, and how we caught it

On screen: the early-Vincennes discovery register, scrolled to the correction.

This is the part I most want librarians to hear.

Twice in two days, web research suggested that recordings from Deleuze's early-1970s seminars were available online. It looked like a finding. Then we downloaded the actual files. They turned out to be text documents only, with no audio at all.

So we wrote a rule into the project: web research turns up hypotheses, not findings. A finding needs contact with the actual object.

Later the archive caught an error of its own. We built a machine-readable dataset of every seminar week. The first thing it did was reveal that a whole run of earlier references to the Spinoza seminar was off by one week. We didn't bury that. We flagged it everywhere it touched and wrote down how to reconcile it.

An archive that can tell you where it was wrong is worth more than one that sounds sure of itself.

Segment five: from lookup to thinking

On screen: the concept index, then the problem index, then one worked example.

Once the texts were solid, we built a layer of indexes on top of them. Here's what each one does:

• A concept index, keyed by nouns: where does "the fold" appear?
• An operational index, keyed by verbs: what does Deleuze do with a concept?
• A register of the philosophers he works through.
• A register of known absences, so the archive can answer "no, he never discusses that" with confidence.
• Threads that follow single concepts across decades, through books and seminars together.

Then came the harder part. We built registers of the deeper structure Deleuze describes in his late book with Guattari, What Is Philosophy?: the planes a problem is laid out on, the figures that thinking passes through, and the recurring kinds of problems.

The real test is what we called re-entry. You bring a problem Deleuze never wrote about, say a school for adolescents, or ecology after the idea of nature, or a political uprising. The archive shows how his machinery would approach it. We ran those as worked examples. Each one exposed gaps, we closed the gaps the same day, and the archive got sharper through use.

Segment six: the Difference and Repetition Engine

On screen: the two governing texts side by side, French left and English right.

The latest stage focuses on one book, Difference and Repetition, from 1968. It's Deleuze's major work.

Here the French governs. Every quotation is given in French first, then in Paul Patton's English translation. Where the two diverge, the French carries the weight, and the divergence goes into a register of false friends. One example: the French distinguishes différentiation, with a t, from différenciation, with a c. That one letter carries a whole argument.

We split the book at its own printed section breaks. There are thirty-seven breaks, which gives forty-five sections. We aligned the two languages paragraph by paragraph, 930 rows in all. We also built a quotation checker. Any phrase in quotation marks has to pass through it before it's filed, and it reports one of three results:

• verbatim
• matches only after normalizing, in which case the phrase is corrected to match the file, never the other way round
• not found, which means it isn't a quotation

The output is a set of questions for dramatization: 113 items across three sections so far. Each one ends on an open question in the book's own forms, like "How much, how, and in what cases?" None of them ends on an answer. That's deliberate, because Deleuze argues that thought starts from problems, not from solutions handed down.

The engine even checks its own grammar. A closing question may not quietly tell the reader what they've already done or what they should conclude. Over three sections that check caught subtler and subtler versions of the same slip, and the rule was sharpened each time. At the most recent re-audit, of the chapter on the image of thought, the review found nothing left to correct.

Segment seven: how the work is actually done

On screen: a Claude session with the project files open, then the same material running in Kimi.

I build everything in Claude. The protocols, the audits, the indexes and the tools all come out of long working sessions there. I've found that the finished instruction sets run better for me in Kimi, so I use both. Claude is the workshop and Kimi is where the finished engine runs.

Every session starts with Claude reading the project's framing document and the last session's log, so nothing has to be rediscovered. Every session that changes a file ends with a written record. No file is ever edited in place. New versions are numbered, and old versions are kept.

Segment eight: why it matters to you

On screen: slow pull back over the whole folder tree.

For librarians and researchers, I'd leave you with three things:

1. Provenance is built into the structure. Every quotation traces to a file, a line, and a fingerprint.
2. Honest limits are part of the collection. Known defects are logged and cited, never quietly fixed.
3. AI can be held to a scholar's standard, if you write that standard down and make the machine check itself against it.

This archive didn't grow through one big leap. It grew through small corrections, each one logged, adding up over time into something that can answer back.