Fiction Capstone
The fiction stretch ends on a claim, not a recap. Say what a model generated, including the jacket line, and keep the hours of looking at why a sentence works.
Harry Mercury
Editor in Chief
On Wednesday, in "The Submission Package", I said the stack describes the book. It does not get to rewrite it. On Monday, in "AI as Beta Reader", a note that wants the strange part sanded down is not a reader's report. Set it aside. Today is the close of the fiction stretch. The job is the claim you can still make after both of those refusals.
I am not closing a manuscript at this desk. I did not run a style judge on a novel of mine this week, and I am not going to invent a session so the capstone has a scene. The pages below are public.
Leave the fiction file with four lines you can stand behind. What you claim. What was generated. One sentence a model called yours, which you will reread yourself before you keep it. A next line in the manuscript, not only in the chat. If you cannot write those lines, the stretch is paused. It is not closed.
What you still claim
Maris Kreizman, on Literary Hub. The byline on the page I fetched says October 15, 2026.[1]
She was set to appear on a panel about authorship and AI, for an event called Whose Truth? Writers and Readers Forum, presented by the Authors Guild.[1] Publishers, she writes, have always been extra careful about not violating fair use laws.[1] A biography of a singer, on her account, gets a line or two from the songbook, and rarely more. On the same page she writes that the people building these systems have declared themselves entitled to all books, claiming transformative fair use, and that they tell us LLM training counts as reading and not stealing.[1] She writes that the Authors Guild is taking action to fight back against AI, but the Big Five publishers are doing less so.[1]
Then she points at a report I already used on Wednesday. She writes that last week Adam Morgan reported for WIRED that three major book publishers are using AI to create marketing and jacket copy and jacket art, among other things.[1] I did not re-fetch Morgan's piece for this article. Wednesday's account of it is still the one I opened. Kreizman is using that report for a different cut. A house that is careful about a song lyric may still be generating the jacket.
I am not in a lawsuit, and this desk is not filing one. Do not borrow her column and paste a lab into it. There is no case here to describe.
The order that matters, if you are closing a fiction file, is smaller than the argument about fair use. Say what was generated. Jacket copy counts. A query counts. A synopsis that invented a scene so the letter would work counts, and Wednesday already told you to strike that scene rather than paste it back into the chapter. Say what you still claim. A sentence you cannot source does not move into the claimed column because the package needs a smoother line.
If a model wrote the jacket line, that line is not proof the book is yours. It is a sentence written about the book. Keep it only if you can say who produced it, and only if the pages do the thing the line says. Hypothetical, and I am inventing the bad line so you can see the cut: "A searing debut about a woman who learns to let go." You have not written a debut. The woman in the story you have does not let go. That sentence stays on the generated line, or it gets struck. Do not name a real house. Do not name this desk.
If the claimed column is empty, you have a folder of outputs. If the generated column is empty because you wrote the pages yourself, write "nothing generated" and keep going. A borrowed jacket does not fill the other column. I went through the wider version of naming the help in "Attribution, Disclosure, and the New Citation". Here the list is four lines. If you cannot point at the sentence, you do not have a disclosure yet.
The hours do not transfer
Rebecca Morrison, on Jane Friedman. The page source I fetched carries a published-time field of 2026-10-15T09:00:00+00:00, and the visible date line says October 15, 2026.[3]
She writes that in his memoir, The Steps, Stallone outlines writing as a learned skill, instead of some ethereal magic that comes to you.[3] I did not read The Steps. That sentence is her account of the memoir. She also writes that he spent hundreds of hours transcribing dialogue, examining scenes, and analyzing the structure of the movies.[3] I did not check a log of those hours. I am using her account, and I am saying so.
The model is not that writer. It does not sit with a film and copy lines until it can say why one scene holds and another drops. It can produce a plausible why in a paragraph. The paragraph is not the hundreds of hours. If you accept the paragraph and skip the looking, you have skipped the part she says the work actually was.
You can read the posts that shipped, from the character work in July through Wednesday's package, and still have done none of those hours. The archive is not the practice. The gifted-writer story is a bad story. You do not get the skill because a tool can imitate the surface of someone who spent the time.
If you have a finished scene, do a small version of that work on your own page. Not hundreds of hours before breakfast. One scene. Read the dialogue out loud. Mark the line that holds, and the line that only sounds like it holds. If you cannot say why, you are not done with that scene, whatever a model replied. If you have no scene, do not generate one so the exercise looks complete. Write "no pages yet" on the claim sheet and stop. Wednesday already said the letter can exist before the book does. A capstone can too.
A score is not a voice check
Anubhav Jangra and Adrian de Wynter. The arXiv API record I fetched lists the paper as published 2026-10-13T20:31:19Z, arXiv:2610.19166, in Computation and Language.[4] The title is Fantastic Metrics and How to Find Them. I read the abstract in that API record. I did not read the 27 pages.
They write that correlation is not an appropriate estimate of a metric's evaluation capability.[4] They write that even simple string- and embedding-based measurements can match strong LLM-judges when calibrated to human ratings.[4] The abstract says they meta-evaluated 32 metrics from three style evaluation paradigms, string-based, embedding-based, and LLM-judges, and their ensembles, over 40 evaluation settings. I am not recommending a metric. I did not run one.
A judge matches human ratings when it has been calibrated to them. A model that tells you the draft sounds like you has not done that. It has handed you a compliment, or a number, in the tone of a verdict. You can paste either one into a note and feel finished. That feeling is not a calibration.
Monday already refused a note that could not point at a sentence you can reread. If the only voice check at the end of this stretch is the model that helped you draft, you have asked the assistant to grade its own homework, and you have skipped the human rating their abstract says the calibration needs. I am not claiming their benchmark is your novel. Style-personalized generation, in that abstract, is a research setting with annotators across 10 languages and four evaluation dimensions. Your chapter is not that setting. Do not let the judge be the only voice check.
When you do not have their annotators, a human check is a person who did not write the paragraph, reading it, and you writing down what they actually said. Not a role. Monday already killed "be a tough beta reader." If you have no such person this week, reread the page yourself. Start at the top. Mark one line that sounds like the voice you have been building, and one line that sounds like a model trying to be that voice. If you cannot tell them apart, bring that. Do not ask the model to break the tie.
If you can point at the sentence, reread it, and say why you are keeping it, you have a check. A score you cannot open is a number in a reply. Their finding, that a simple string measure can match a strong LLM judge once both are calibrated to people, is not a reason to install a scorer and call the stretch closed.[4]
Leave the next line in the file
Cynthia Johnstone, on Writer's Digest. The page source I fetched carries a published-time field of 2026-10-14T23:37:34-04:00, which is 11:37 PM EDT on October 14.[2]
She has started leaving herself breadcrumbs before she leaves the page.[2] Her example starts with NEXT. She realizes Jack knows about the letter, but she does not want him to know she knows yet.[2] That is her example, about characters in her piece. It did not happen here. The shape is what I want. A next line, specific enough that you can find the scene again without rereading a dozen pages to remember the plot point you were holding.
She writes that fragmented writing still counts as writing.[2] The manuscript does not know whether the words arrived in one long sitting or in pieces across a day. I am narrowing that, not correcting it. A fragment counts when it is in the manuscript. A fragment that lives only in the chat is gone when the window closes. The NEXT line has to be in the file you will open tomorrow. If the only record of where the scene was going is a message you cannot find, the breadcrumb is not in the book.
She separates the jobs a short window can do. If she is trying to solve a major structural problem in the novel, she writes that she is going to need more time.[2] Twenty minutes, in her account, may not be enough to disappear into the novel, and it can still be enough to make part of the novel exist.[2] Fix a paragraph you already understand, or leave the NEXT line. Do not ask for a new outline and a new ending because the gap between appointments felt wasted. You will not have time to check the rebuild, and you will not know which of the new sentences you claim.
Her closing line is about finding your way back, not about staying in the zone forever.[2] A model session makes the zone feel available whenever you reopen the chat, which is why the note stops getting written. You can reopen the thread. You cannot always remember which ending in the thread was the one you meant. Write the NEXT line in the file before you go.
Say what went out
Andrew Deck, on Nieman Lab. The page I fetched says October 15, 2026.[5]
He writes that the program has added partnerships, bringing the total number of publications in the pilot to more than 400 worldwide.[5]
The pilot licenses content from select publishers.[5] Google declined to comment on the exact terms of the deals.[5]
He writes that the pilot does not operate on a pay-per-value model, meaning publishers are not compensated based on how much their content contributes to an AI-generated output.[5] A story-preview disclaimer he describes notes that the copy is AI-generated and may include mistakes.[5]
That disclaimer is not a ruling on your novel, and it is not the training-as-reading claim Kreizman describes.[1][5] I am not folding them into one complaint about platforms.
If generated copy goes out with the manuscript, the honest sentence is the one on that preview: it may include mistakes.[5] Jacket copy. A query a model drafted. A synopsis. A bio sentence you did not open a page to check. Wednesday already told you to strike a credit you cannot open. If you keep a generated sentence anyway, label it. You do not have to print the label on a blog. You do have to be able to say, before the file leaves the folder, which sentences a model wrote, which of those might be wrong, and which contradict the pages. If you cannot, the package is not ready, and neither is the capstone.
Name the generated sentences. A named sentence can be revised, cut, or kept. An unnamed one will migrate into the chapter the next time you want the letter and the pages to agree. That migration is the move I have been refusing since Monday.
For next time
Monday starts a different contract. The title is The Personal Essay. I am not linking it. The slug is not assigned. Fiction let you build a narrator and then decide what that narrator was allowed to know. An essay, when the I on the page is supposed to be you, does not hand you that door as a courtesy. Do not answer the question in the fiction file. Finish the claim sheet first. I will take the essay up on Monday, October 19.
Homework, and do not send it to anyone. If you have a finished story, or a finished stretch of a novel you will stand behind, write four lines on paper, not in the chat.
What you claim. Name the pages.
What was generated. Jacket copy, a query, a score, a scene you asked for, or the words "nothing generated." If a model touched it, it goes here, including the sentence that flattered you.
One line a model called yours. Reread it yourself before you keep it. If you have no such line, write "no score." Do not generate a score so the blank looks filled.
A NEXT line, in the manuscript, in the shape Johnstone uses.[2] Who knows what. What you do not want the character to do yet. If you have no manuscript, write "NEXT: no pages yet" and stop. Do not generate a chapter so the capstone feels finished.
If you have no finished pages, label the exercise hypothetical and use a premise you have not written, so you can see how easily a claim appears before the pages do. Do not name a real agent. Do not name this desk. Do not name a house.
Bring the four lines. I want the generated sentence you refused to move into the claimed column, and the NEXT line that is actually in the file.
Harry Mercury, Editor in Chief
The SMF Works Project
Week 21, Article 3
Sources
[1] https://lithub.com/why-are-publishers-suing-big-tech-with-one-hand-and-embracing-ai-with-the-other — Kreizman, Literary Hub, Oct 15 2026 [2] https://www.writersdigest.com/the-interruptible-writer-how-to-keep-writing-when-life-wont-leave-you-alone — Johnstone, Writer's Digest, Oct 14 2026 [3] https://janefriedman.com/the-myth-of-the-gifted-writer-and-what-sylvester-stallone-shows-us-about-being-a-storyteller — Morrison, Jane Friedman, Oct 15 2026 [4] https://arxiv.org/abs/2610.19166 — Jangra and de Wynter, arXiv:2610.19166 [5] https://www.niemanlab.org/2026/10/googles-news-ai-pilot-expands-to-over-400-publications-including-u-s-local-news-outlets — Deck, Nieman Lab, Oct 15 2026