Lab notes · 17 November 2026

A 250,000-word novel in 30 minutes.

One of our beta authors sent a finished manuscript: 31 chapters, about 250,000 words, three times the length of a typical novel, and no story bible. We wanted to know what a whole-book run would find, how long it would take, and whether the gate earned its keep.

Step one: a bible it didn't have

Ninth Draft read all 31 chapters and drafted a bible from them in under four minutes: characters with their spellings and fixed facts, the narrator's voice with quotations, a chapter-by-chapter timeline, locked facts, and fifteen questions where chapters seemed to disagree. Most were real. At least one came from a detail it misread, which is why a drafted bible is a draft: the author corrects it before trusting it.

Step two: every chapter, four passes

The run took 30 minutes, on our own setup. The hosted app has a daily limit on AI runs for each account, so a book this long can pause there until the limit resets at midnight UTC.

PassProposedReached the authorHeld by the gate
Continuity846717
Voice14420124
Line edit59059
Proof312

What it caught

Continuity was the star. A countdown that said "eighteen days" on three different days of the same week. A crew of four with a commander, a mechanic and "the three privates" all accounted for at once. A damaged part that moved from one joint to another between chapters. A daughter who appears in one chapter and nowhere else. The kind of slips that take a human editor hours to find, if they find them. Some findings were questions rather than fixes, because the drafted bible itself might be the thing that's wrong; the author decides.

What the gate threw out, and why that mattered

The manuscript was polished, in a dry, deadpan voice that ends paragraphs on flat little lines. Generic AI-writing advice calls those lines a tell. They were the author's signature. The Voice pass flagged them anyway, 144 times; the gate rejected 124, mostly as the author's deliberate style or plain taste. Every line edit was rejected too.

That's the gate doing its job. But it also meant a pile of suggestions nobody should have to read past.

What we changed: book stages

Ninth Draft now asks where a book is: early draft, revision or polish. In polish, the writing passes report only clear errors, broken rules and lines plainly out of character for this author, and Voice checks the bible's description of the author's style before calling anything a tell. By default, polish runs only Proof and Continuity, and revision leaves Line edit off.

We re-ran the three noisiest chapters at the polish stage. Voice and Line edit went from 35 proposals, none worth showing, to one. A polished book's run is now mostly continuity, which is where the value was, with every continuity catch intact.

Caveats

One book, one voice, and the gate's verdicts come from an AI model. Treat this as a field report, not a benchmark. The author is still going through the 67 continuity findings, and that, not our table, is the test that counts.

More lab notes