The Author Engine

The Author Engine is an operating architecture for producing long-form work with artificial intelligence while keeping the author, rather than the model, in command. It brings drafting, structural review, prose analysis, source verification, release control, and cryptographic attestation of the work itself into one governed production system. The proposition is straightforward: AI contributes speed and analytical reach, but the manuscript advances only through defined evidence, checks, and human-authorised decisions. The result is not simply generated text - it is a controlled body of work whose development can be inspected, challenged, and traced.

The system addresses the central weakness of AI-assisted long-form writing: capability without dependable continuity or control. A model can produce persuasive prose while drifting from the author's intent, contradicting earlier chapters, flattening an established voice, or presenting a plausible citation that does not support the claim attached to it. Over a substantial manuscript, these failures compound. The Author Engine treats continuity, prose quality, factual accuracy, and citation integrity as managed production requirements - not matters to be repaired informally at the end.

Its strongest credibility claim is not perfection. It is the documented paper trail of the system catching itself being wrong, repeatedly, and hardening as a result - roughly 120 adversarial findings across live operation. In one documented case, a reviewer attacked a placeholder rather than the actual text, creating a phantom review round. In another, a reviewer cited quotations that appeared nowhere in the document. Both failure modes are now caught mechanically before a finding is even weighed. When a language model's unaided estimate of a chapter's sentence-length distribution returned three mutually contradictory figures, cadence measurement moved to a deterministic script - never an AI's own guess.

This is not a laboratory concept waiting for a first use case: four books currently run this production architecture - three complete, one in production - alongside the tooling project that documents the system itself.

The six engines

• IE

  Initiation Engine - project setup

  Owns everything between naming a project and clearing the first material dispatch: folder-tree provisioning, instrument instantiation, and a hard readiness gate. "First engine provisions all" - and every engine that binds afterward independently re-verifies the tree rather than trusting the first pass. The readiness gate blocks all material work until every required configuration field is present or explicitly declared absent.

• DE

  Design Engine - forging the specification

  Where a book's mechanisms, world-rules, and logic are argued out, attacked, ruled on by the author, and locked before a word of prose is written: a specification in which every mechanism is argued rather than asserted, every settled number is on the record with its reasoning, and every deliberate gap is named. Runs a seven-step adjudicated design cycle with invited red-team attacks; any reviewer finding built on a quotation that does not match the real text word-for-word is discarded on the spot.

• PE

  Prose Engine - writing the book

  Turns the ratified specification into chapter prose through seven author-held gates, including a mandatory direct read-and-edit pass in Word. Authorship and judgement are strictly separated: the controller never writes book prose, a dedicated drafting lane is the sole drafter, and an independent cold-reader observes without rewriting. No lane ever adjudicates prose it authored.

• RE

  Regeneration Engine - cadence and craft remediation

  A standing, discretionary loop for rhythmic and mechanical correction of already-ratified text. "RE never authors" - every word it touches was ratified before RE opened it, and nothing it produces becomes canon until it clears review again. A craft-language firewall keeps numeric targets out of drafting dispatches entirely.

• VE

  Verification Engine - citation and claim integrity

  The manuscript-wide, last-stage check. Its central question: does a cited source actually say what the manuscript claims it says? A ten-stage ladder scans claim-citation hazards, checks the source's actual content against the claim, rectifies every flag against a closed disposition catalogue, reflows the manuscript, re-verifies after the edits, and closes with blind-first whole-book citation-integrity and originality audits before the submission-build lock releases.

• SE

  Signature Engine - attestation and chain of title

  Answers the two questions a publisher's lawyer asks and no amount of good prose can settle: what existed, when, and whose it is. Named artefacts are fixed by SHA-256 manifest, chained by predecessor hash, and stamped with a timestamp token from a public authority outside this press's control, held in three replicas. That establishes what a file was and when it existed. A second half, in build, determines authorship: it reconstructs a manuscript's paragraph-level history across every dated state, tests each surviving passage against the full record of AI-assisted conversation behind the book, and separates what the author composed from what remains machine-generated. The output is an evidentiary record, not a legal opinion.

How the system argues with itself

The Design Engine is the system's one deliberately adversarial engine: its reviewers are told to attack the mechanism under review, not observe it. The Prose and Verification engines use a lighter pattern - cold-read observation rather than invited attack - with the final whole-book audits run blind-first by one lane and then reconciled by another. And the adversarial culture runs all the way up: there is exactly one place where an AI lane is explicitly licensed to push back on the human author, stated as a two-way street - pushback is welcome where the argument warrants, and a good argument is expected to move the author's position.

Two red-team failures are worth telling as evidence rather than war stories. In the phantom-round incident, a review dispatch carried placeholder text instead of the actual document - and the reviewer returned three confident, well-structured findings against a document it had never read. In the fabricated-anchor incident, a reviewer anchored its criticism to quotations that existed nowhere in the text. Both were caught by the same discipline, now mechanically enforced: any finding that opens with a quotation not verified word-for-word against the real text is discarded before its merits are even considered.

Working at scale

Within a single project, up to six parallel controller lanes can run concurrently against different chapters - six being the proven complement - without cross-lane state corruption. Across the estate, four books and the tooling project itself run the identical production architecture, each operating independently with its own instruments and cross-referenced through a standing messaging layer rather than merged into one shared state. Anything an AI is tempted to estimate is measured instead: sentence cadence, file integrity, and folder structure are all verified by deterministic scripts, never by a model's own guess.

Failures that built the system

Every integrity rule in the engine traces to a specific, documented incident rather than abstract best practice. A reviewer attacking placeholder text produced the verbatim-anchor rule. A controller running a full remediation loop without reopening the governing instrument produced the stage-transition re-read discipline. A project configuration missing one required field - uncaught until it nearly reached a prose commission - produced the hard readiness gate. Two files sharing both an identical name and an identical byte size produced the rule that deletions match by unique file identity only. The system's record is not that it never fails; it is that every failure becomes a mechanical check that prevents the next one.

Chain of title

Most systems for writing with AI can tell you what they produced. Very few can tell you, afterwards and to an outsider's satisfaction, who wrote which parts of it - and that is the question that decides whether a work can be registered, warranted to a publisher, or sold with clean rights. The Signature Engine exists because attestation and authorship are one question with two halves. Fixing a file's identity and date proves that a manuscript existed in a given form at a given moment. Reconstructing its composition history proves who made it. Neither half is worth much on its own.

The method is deliberately reproducible rather than authoritative. It is published in full, it runs on evidence the author already holds, and it is designed so that a sceptical outside party running the same procedure on the same material reaches the same answer. That has been demonstrated: the measurements underlying this press's own first case were independently re-executed by a different vendor's model, working from the published method and pulling the source files itself, and reproduced every figure exactly. A method a rival can run is a method a rival cannot easily dismiss.

Used forward, from the first day of a project, the engine removes the problem rather than measuring it: every save is stamped as it happens, so the record of who wrote what is contemporaneous rather than reconstructed. Used backwards, on a finished manuscript, it is forensic, and it degrades with the quality of the record that survives. Both modes report their own uncertainty on the face of the result. A figure that has not been independently checked says so.

Governance

A small set of integrity gates holds across every engine, each born of a specific, documented failure rather than abstract best practice: nothing is deleted without an archived, verified twin already confirmed in place; nothing is treated as live until independently confirmed in its home location; and no workflow stage closes on memory - only on a fresh re-read performed after the stage actually closed.

Authorship and rights

Every word of book prose is written under a specification the author has personally ratified at every gate, and every chapter closes only on the author's explicit sign-off. The author is positioned at every gate and at multiple points of creation and review precisely so the completed work is intellectually owned by the author. The engine's role separation makes open, role-precise disclosure of AI participation possible - designed against the human-control standard in current US Copyright Office guidance (2025) - and completed manuscripts are additionally screened through external, non-AI similarity tooling before release. The Signature Engine turns that disclosure from an assertion into a record: what is claimed, what is disclaimed, and the dated evidence behind both.