The method

A language model is an unreliable industrial component.

Most tools call a model and display the answer. We treat a model call the way a factory treats a machine on a line. Three properties, all live today.

The three properties

Every editing pass is a job with a record.

Each pass a house starts is registered before it runs, carries a timeout, and ends in an explicit terminal state — ready, failed or timed out. A pass that never finishes is detected and recorded rather than quietly disappearing.

One door for the editors.

Every editorial read — Alex, Sam and Jordan — passes through one component that records the model, token counts and cost of each call.

A file we cannot read properly is refused.

A manuscript that fails validation is refused with the reason stated, never loaded in a degraded state and never silently repaired. A completed stage shows whether it was completed by a person or by the system.

Measured, not modelled

Figures from real manuscript reads in production, October 2026: two books, at 47,000 and 64,000 words. Outside that range a timing is an estimate, and we will say so.

31m 44s

47,000-word read

32m 55s

64,000-word read

6

model calls per read, each recorded

82

chapters, longest run

Turnaround is close to flat with manuscript length — a 36% larger book took 4% longer, because the five analyses run concurrently and wall-clock time is dominated by model latency, not word count. A read does not depend on anyone’s browser: the job record exists server-side, and completion shows on the house’s Books list.

What this buys you

Every one of those properties exists so that a result you disagree with can be taken apart. If an editor at your house says a finding is wrong, there is a record of which call produced it, and against which text. A tool that cannot be audited cannot be argued with, and an editorial opinion you cannot argue with is not much use.

Put a manuscript through it

The fastest way to judge us is the product’s own: send one manuscript through the read and compare the result with your editors’ view of the same book. Write to us — a person reads every enquiry — and we will also send the technical specification: what the system does, how it is built, and what it does not yet have. A specification that contains only good news is a brochure.

publishers@authorslab.ai