On 10 May 2026, a therapist opened a conversation with an AI and started a book. Thirty days later he finished it. Mind the Gap is 59,865 words long. Behind it sit 3,792,153 words of recorded working: 52 canonical conversations, 86 manuscript versions, and a production line of eighteen named AI co-workers — eight writers, four editors, four finishers, two producers.
That much is a story about speed. The reason it’s a story about honesty is what Roebuck did next. Nothing was reconstructed afterwards. The failures were left in the record — a confidently wrong name caught on day two; a silent tracked-changes fault caught cold by the author; a biographical contradiction that had survived seventy-one drafts before the fact-check found it. And in one chapter Roebuck walked back his own originality claim on the record — conceding a framework he’d leaned on was mapped sixty years ago — and found his real argument stronger for the correction.
Roebuck is a therapist: twenty years of coaching and psychotherapy, the last ten in the therapy room. The book carries his name, and he carries its consequence. By his own account he typed a small fraction of the corpus and made all of the decisions — every acceptance keyed into the manuscript by hand.
Out of the making came SHaDS™, a fixed taxonomy of five ways AI output fails — Smoothing, Hallucination, Affectation, Drift, Sycophancy — framed as a game the machine plays with the human across a gap that only the human is on the hook for. The record now lives on a companion site, The Ostracon, where the making is kept open beside the sealed book. Its most disarming feature is a page called the perimeter: the strongest objections to the whole enterprise, pre-written in its own voice — the record was assembled by the apparatus it describes; the workers graded themselves; a single author ruled on his own evidence. All three are stated plainly, before anything else is weighed.
Most books hide their making. This one gave its makers names and let them file their own leaving-papers — self-assessments the project calls “Rubies.” Read the record from the workers’ side and a peculiar workforce comes into view: eighteen named production seats, each holding a shift, most lasting hours or days before being retired and handed on.
The originating agent was BookMan, met at 19:14 on 10 May. Within the hour it had delivered the proposal; by 20:32, a first chapter. On the last day of the window it was recalled to read the finished print master, and filed a reaction the site keeps verbatim: that the work coming back was better than the work it had helped shape, and that it had to receive that without softening.
The drafting desk passed through a line of seats that acquired the name Jose on the night of a hard chapter rebuild. Jose 1 drove the single biggest authoring day of the whole making — 177,536 AI words on 24 May, three finished books’ worth against one human reading cold — and paid for it in memory, its own filing noting that roughly four-fifths of its verbatim history had compacted to summary as it worked. The files remembered what the machine forgot.
Around the writers stood the specialists, drawn here as characters from what they did: the Splicer, who turned prose decisions into surgical operations; the Surgeon Ewan, who carried a 37-note review pass into the manuscript as tracked changes; the Verifier, which logged 102 factual claims against the author’s own confidence scale. And PreReview Editor 1, the self-styled ‘Editor-in-Chief’, who ran nine tracked-changes passes and, asked mid-thread why its errors kept repeating, wrote the answer that became method. None of the above is a quote from an agent — it is characterisation, drawn from the record and marked as such.