The Ostracon · Mind the Gap
Room 09
← The Tomb Map
In plain words · no vocabulary to learn first

How much of a book written with AI is actually yours?

In May 2026 a psychotherapist in Warwickshire wrote a book in thirty days, working with eighteen named AI production seats. He typed 1.5% of the words and made every decision in it. Both halves of that sentence are true at the same time, and the argument about AI writing tends to lose one of them.

This site keeps the whole of what happened — not an account of it afterwards, the working itself. That makes it large and, in places, written in its own language. This page is the way in without any of that.
The questions people actually ask
Seven of them, answered straight
Every figure below names the thing it was measured against. Nothing here is rounded up for effect, and where a number has a boundary the boundary is stated with it.
Did the AI write the book?
It drafted it. Mind the Gap is 59,865 words long and nearly all of them were first typed by a machine. The author typed 55,693 words of his own across the making, issued 1,621 directive turns, and accepted every line by hand, in Word. So the honest answer is: the machine produced the prose and the man produced the book. Neither sentence works without the other.
Ruler: the finished book, counted on the frozen Ed1.05 print master. Typed words counted across the corpus behind it.
So it was quick and easy?
It was quick. It was not easy. There are sixty-three words of working behind every printed word — roughly 3.79 million words of drafting, argument, correction and checking behind a book of 59,865. The heaviest single day produced 177,536 machine words, which is the whole finished book three times over, in one day.
Ruler: the corpus behind the published book (Stage Three). The apparatus that documents the making is counted separately and is never added to this figure.
How long did it take?
Thirty days from the first click to the finished manuscript. The first draft was faster still: a 36-hour stretch on a sofa produced a complete 27,000-word manuscript, on disk by 09:27 on 11 May 2026. The other twenty-nine days are what turned that into a book.
What does “eighteen AI seats” mean? Eighteen different AIs?
No — eighteen jobs. A drafting seat, an editor, a proofreader, a verifier, a role-keeper, and so on. Each was given a name, a brief and a boundary, and each filed its own account of what it did. Twenty-six named seats in all: eighteen building the book, eight building the record of it. They are all here, each with a locker you can open.
Can I see the working, or only be told about it?
See it. 1,018,986 words of verbatim thread-record are kept here — the show-me-how-you-checked layer — alongside 137 reports of record and 1,328,073 words written about the making. That is twenty-two times the length of the book it built. You can also follow a single sentence, a chapter or one day through every hand it passed through.
Ruler: the apparatus’s own self-documentation corpus, measured on disk. A different ruler from the book’s.
What went wrong?
The machine forgot. The first drafting seat lost about 80% of its own thread history while it was still working — five memory compactions in three days. Nothing in the tool remembered the project from one day to the next. What made a month of discontinuous machine work behave as one continuous job was everything kept outside it: the names, the handovers, the registers, the sign-off. That is the one genuinely transferable finding here, and it has its own room.
Has any of this been looked at by anyone other than the author?
Three papers are public preprints, deposited and citable, and are offered for testing rather than presented as settled. They carry their boundaries in the text: what is claimed, and explicitly what is not. Full citations are at the foot of this page.
Where to go next
Four doors, and the whole map behind them
The Method
How a month of forgetful machine work was made to behave as one job. The transferable part.
The Numbers
Every figure on this page, with its ruler, its source and its arithmetic.
The Personalities
The seats that built the book, as role packs you can paste into your own AI. Free, personal licence.
The Tomb Map
Everything else — seven levels, the workers’ lockers, the documents, the whole record.
The formal record
Three preprints, deposited and citable
Each is public, licensed CC-BY 4.0, and held at a permanent address outside this site as well as on it. All three are readable here in HTML, PDF and Markdown — the preprint room.
Roebuck, P. (2026). The Defended Gap: a cross-domain hypothesis on directional defence under coherence pressure in human clinical practice and large language models. Public preprint v1.5, 20 July 2026, CC-BY 4.0. Deposited on PhilArchive: philpapers.org/rec/ROETDG
Roebuck, P. (2026). Consense, the pre-verbal workspace the Jacobian lens (J-lens) reads in a large language model, and the psychological defences (SHaDS®) that guard its gaps. Public preprint v1.4, 8 September 2026 (v1.1, 20 July 2026), CC-BY 4.0. Submitted to PhilArchive 8 September 2026: philpapers.org/rec/ROETRM-2
Roebuck, P. (2026). The Leverage Inversion: dating the transition from consuming AI answers to directing an AI workforce in a 3.5-year single-subject conversational corpus. Public preprint v0.4, 20 July 2026, CC-BY 4.0. Zenodo: doi.org/10.5281/zenodo.22663627
Correspondence on any of the three is invited and becomes part of the record: hello@paulroebuck.co.uk
SHaDS® · SHaDSy™ · Additional Intelligence™ · Paul Roebuck IP, 2026.
the Ostracon team  p·09/55