Mind the Gap · the making record · 10 May – 9 June 2026
Contemporaneous documentation of human creative control in an AI-assisted work: a complete instance
The Office’s Part 2 Report (Jan. 2025) settled that AI-assisted works are registrable where a human exercises meaningful creative control over the expressive elements; Thaler v Perlmutter (No. 25-449, cert. denied Mar. 2, 2026) fixed the human-authorship floor while leaving open how much human involvement suffices. The live question — posed directly in Allen v Perlmutter (D. Colo., pending) — is evidentiary: what does demonstrating control look like in practice? As of July 2026, cross-motions for summary judgment in Allen are fully briefed and await decision.
This record is one worked answer, preserved rather than argued: the complete working corpus of a 59,865-word trade book produced with AI assistance over thirty days — every correction logged, every voice-preservation decision visible, every version difference tracked, captured at the moment of creation and not reconstructed for any proceeding. Where A Single Piece of American Cheese registered on 35 documented rounds of iterative modification, and the Zarya decision preserved authorship at the selection-coordination-arrangement layer, this record documents that layer continuously, at corpus scale, with the human and machine contributions separable throughout.
And the posture here is not Allen’s. There, the expressive output is wholly machine-generated and the human case rests on the prompts; the Office’s briefed reply is that prompts are instructions, that the system determines the final expression, and that effort and iteration do not substitute for authorship. Here the contributions are separable at every step: 55,693 words typed by the author, 1,621 directive turns steering selection, coordination and arrangement across 86 versions, a human reviewer’s 37 page-keyed items carried in as tracked changes — every acceptance the author’s, in Word, by hand. The claim is never the volume of the iteration. The claim is a documented human hand at the layer Zarya preserved, held continuously, at corpus scale.
The author typed 1.5% of the corpus — and made 100% of the decisions.
Of the 1,621 directive turns, 346 were ratifications and 177 were corrections or outright refusals — roughly one human turn in three was judgement passed on machine output rather than new instruction. Every turn was classified by hand after an automated pass was tested against a blind sample and rejected as inadequate; the labels validated 100/100 against two held-out samples, and the count reconciles to the register exactly.
The record holds 52 documented rejection events — a floor twice over: the count covers the chat channel only, hand-validated from every candidate, and the author’s accept-or-reject rulings on tracked changes in Word are a second channel the transcripts cannot see. Among them: the machine’s proposed ending, killed for the author’s own typed closing lines — the book’s last words are his.
And the typed words that survive are the load-bearing ones. Of the author’s 55,693 typed words, 2,155 — 3.9% — survive verbatim or near-verbatim in the print master, and the survivors are the title, the dedication, the epigraph, the disclosure, the aphorism the book is built on — each surviving passage mapped to its printed page. Small by volume; the spine of the book by weight. Survival is not proof of origination direction — dictation and quoting-back-while-editing both count, and both are authorial acts.
The silent channel is measured too. 120 of the 257 retained manuscript files still carry the machine’s tracked changes frozen in situ — the proposals preserved beside the author’s decided resaves, often minutes apart, the channel unbroken from the first version to the print files. Of fifty measured edit rounds, thirty-four were accepted whole, fourteen mixed by hand, and two rejected entire — in Word, leaving no chat trace. At one of those, the author then typed the back-cover page himself: thirty-three tracked insertions under his own name.