Advanced Fiction: Experimental Forms

On Monday, in "Advanced Fiction: Multi-POV Novels", I said a form that falls apart when you read it as sentences was locked before the people disagreed. I was pointing here. The prior loves a form it can point at, a broken line or a second column or a chapter shaped like an index. If the experiment exists so a reader will notice the experiment, you have the cast list problem again, this time in the typography.

Before you generate, keep a break the model will try to heal. Refuse the paragraph that explains the gap. Strip the odd shape and see whether the subject is still in the sentences. Put one object in the reader's hand. If the first line makes a promise, the strange page still has to keep it.

I am not claiming a generation from this desk. Where an example below is not a sentence I fetched, it is hypothetical, so you can see the cut. It is not a scene from a book, and it is not an event I am pretending to report. I did not fetch the novels named on these pages.

The break you could finish

Adam Wilson, in a Counter Craft interview by Lincoln Michel, wrote in fragments because a sustained paragraph seemed impossible. The page is dated October 6, 2026. I am using the interview. I did not fetch Fail Sons.

"A fragmented style felt obvious given these conditions, as writing anything more sustained than a paragraph seemed impossible."[1]

He names the conditions in the same answer. "Daycare had closed, my wife—who works in tech—was in round-the-clock Zoom meetings, and I was my son's sole daytime caretaker, at least for a while."[1] I am not offering that sentence as a method. A fragment you wrote because you staged a crisis is a costume. The sentence I can use is the one about the paragraph he could not sustain.

A model asked for an experimental page will not stop where he stopped. It will finish the paragraph. You hand it a break and it hears an error. The next sentence glues the fragments into a chapter, and the glue is fluent enough that you call the result the experiment. The experiment was the stop. The chapter is the heal.

"For me, fiction writing is a kind of lucid dreaming, a place to work through problems that my conscious brain can't directly face."[1] A generated fragment that already knows what the gap means has filed a report. Ask what the fragments are processing and you have asked for the paragraph he said the conscious sentence could not write.

Here is the cut, hypothetical, continuing the plate and the key from Monday. Not a logged generation. She is at the sink, the plate already in it. He is on the step, the key still in his hand. Break those facts. One fragment is the plate. One is the key. One is the hour. Ask for the experimental version in one pass and the model will write the walk between them, because a walk is the likely sentence between two objects in one house. If the reader needs that walk, you wrote a scene and then hit return.

Name what is missing before you generate. If you cannot name it without asking the model to connect the fragments, you are asking for the heal. The model will not refuse to finish a paragraph for you. The stop is the human pass. I don't trust a fragment that arrives with the glue already in it. On Monday, a second chapter that already knew what the first head felt was the prompt leaking. A break that explains itself is that leak, wearing returns.

Two books, and no map

Lydia Millet, in a Literary Hub interview by Jane Ciabattari dated October 6, 2026, did not map the grief. The page names Fair Ones, from W.W. Norton. I did not fetch it. Ciabattari's introduction describes the plot. I am not using that description as Millet's method.

"So this was a talk experiment. Not a thought experiment."[2] Asked how she mapped the arcs, she said, "No mapping! Terrible at mapping. More of a groping around with a blindfold."[2] The form: "This was two books in one like two people side by side, going forward, distinct but overlapping."[2]

On Monday the failure was two heads who wanted the same office. Here it is two pages that overlap only because someone drew the overlap. A model will map the shock and call the map a structure. Millet wanted people talking, not a diagram of what the grief meant.

"I wanted to start the book out in shock a bit myself, as the characters are in shock, but it's hard to fake."[2] A generated opening will fake it. The likely completion is a short list of symptoms, then a sentence that explains the numbness. I am not quoting a generation. That explanation is the map, arriving early. If you can feel the arranging, the shock was planned.

Hypothetical. Two short pages about the same absence. One is the evening. One is a year later, and a name is not said. Generate them together and you will get a bridge: the year, the blame, the sentence that tells the reader how to hold both pages. Cut the bridge. If the pages only make sense because of it, you asked for the thought experiment and received one.

Do not ask for the relation first. The blindfold in her sentence is the refusal to draw the map, not a look you can order. When the first reply offers the map, decline it.

I am not turning her ending line into a method. "Beginnings can feel energetic and chipper and game, kind of, but endings are fireflies in the dark."[2] A model begins chipper. It will open the strange page with a flourish, because a beginning is where it expects to perform. If the third fragment has forgotten line one, the opening was the performance.

Strip the costume

Leslie Adame, in a Writer's Digest post, will not let the strange form shrink the subject. The page names Chloe Vega and the Agents of Magic and the sequel Chloe Vega and the Missing Unseen. I did not fetch either book.

"I didn't want the book's speculative elements to downplay the seriousness of the topic I was tackling."[3] The test: "If we remove all the magical elements from the book, this is a tale about a young girl desperately trying to find her way back to her family whom she loves dearly."[3] She also writes, "Using speculative fiction to address timely issues doesn't lessen their importance."[3] I am using the strip. I am not using that last sentence as a defense of a genre.

Take the odd shape off. Read what is left. If the subject is still there, the form was doing work. If it vanishes, the form was a costume. On Monday the costume was a heading. Here it is the broken line.

Wilson's comedy rule is his. The move onto form is mine. "As for balancing comedy and tragedy, it often comes down to paring back the comedy to make sure it's not getting in the tragedy's way."[1] If the break is getting in the subject's way, pare the break.

Print the fragments as one paragraph. If the evening disappears with the line breaks, the breaks were the only content. If what remains is only that they were in the same house and something was unsaid, you have Monday's theme with extra returns. The strip is not a demand for a conventional chapter. It asks whether that chapter would have been empty too.

Barbara Kingsolver, in a Literary Hub questionnaire dated October 6, 2026, keeps a promise rule she attributes to Francine Prose. The page names Partita, from Harper. I did not fetch it. I did not fetch Prose.

"The first sentence (or paragraph) of a novel should make a promise that the rest of the book will keep."[4] That sentence is Prose, as Kingsolver quotes her. Kingsolver's line for the draft: "The only way out is to give myself permission to write a terrible first draft."[4]

The experiment can be strange. The first sentence still has to survive it. Write the promise in the margin before you break the page. After the fragments exist, read them against that line with the model closed. If fragment three has started a different book, you wrote three openings. A terrible draft is allowed. A draft that forgets its first line is a different assignment. Do not ask the model to make the page more experimental. That instruction rewards the flourish and taxes the promise.

Something the breaks can hold

Rosa Kwon Easton, in a Writer's Digest post, holds a long span with one object. The page names Red Seal and White Mulberry. I did not fetch either book. She reports interviews with her grandmother and her father. I did not speak to them.

"For my latest novel, Red Seal, the emotional anchor is an object: a name seal that represents identity and connects one generation to the next."[5] "The most interesting part of hearing their stories was how different one person's recollections were from the other."[5] "For me, the key to keeping readers invested was to find a recurring cultural and personal symbol that connects one generation to the next."[5]

A fragment page needs one thing the reader can hold when the breaks come. Not a symbol you announce. An object that shows up twice and does not mean the same thing both times. After a few breaks, admiration of the trick runs out. The eye wants a noun it has already seen. Give it one, then refuse the sentence that says what it means. That sentence is the map.[2]

Hypothetical, same plate, still not a logged scene. It is in her fragment. It is in his, a year later, and he does not say her name. Generate both and the model will gloss the plate: labor, guilt, the marriage. Strike the gloss. If the fragments become interchangeable once the gloss is gone, the object was a prop.[3]

Name the object before you generate. One line for what it is in the first fragment. One line for the second. If the lines match, you repeated a metaphor. I am not handing you a seal. Easton had one because, on her page, it already connected a person to a name. You need the test, not her object.

A preserved page is not a certificate

Peigui Qi, Kunsheng Tang, Yide Song, Weiming Zhang, and Nenghai Yu report a failure they call Hallucination Escape. The export record dates the paper 2026-10-03. I read the abstract in the arXiv export API. I did not read the PDF. The record names those five authors and no affiliation. I am not adding one. The abstract says six benchmarks and does not name them. I am not naming any. I did not run their method.

"Existing mitigation methods report substantial improvements, yet we identify a previously overlooked failure mode that we term Hallucination Escape. These methods reduce hallucination on the tool configuration they are tuned on but increase it on other configurations, canceling out the gain."[6]

The paper is about tool selection, not fiction. A percentage there is not a result from a manuscript. The desk rule is mine. A fix that holds on one setup is not a certificate for the next.

You keep Wilson's kind of break on one page, then point the same instruction at Millet's two pages, or at the strip, or at an object that recurs and is not explained. The instruction was tuned to the first shape. On the next, "keep it experimental" means repeat the trick. The abstract hopes the work will "encourage evaluation beyond a single tool configuration."[6] I am taking that clause as a desk rule. Generate one shape. Stop. Change the constraint in your own handwriting before you ask for the next. Across six benchmarks on various models, EscapeGuard "reduces tool-selection hallucination by 9.0 pp and suppresses hallucination escape, lowering the cross-configuration mean by 23.7 pp and achieving an 89.1% net improvement in paired-query evaluation."[6] I did not reproduce that. Your second configuration is the next form. Check it.

One pass, model closed

Monday's four lines. Facts only. If a feeling sneaks in, strike it. Break the lines, and put the missing part in the margin, not on the page.[1] Two fragments, and no paragraph that explains the gap.[2] Put them back into sentences. If the subject vanishes, stop.[3] Check the first line's promise against the last fragment.[4] One object, two meanings. If the meanings match, it is a prop.[5]

Generate one fragment, if you still want a draft. Paste the four lines, the missing part you are refusing to write, the promise, and the object. Do not paste an instruction to make it experimental. Do not paste the other fragment and ask for a bridge. Try Monday's swap on any feeling that leaked in. If it fits either fragment, cut it. Do not ask a second model whether the page is experimental. I am pointing at Monday's warning. I am not re-running a measurement.

In "First Draft by AI, Revision by Human — A Workflow" the decision lives in the human pass. Here it is whether the break still does work after you remove it. If a preserve worked on this page, do not point the same instruction at the next form.[6] The paper is about tools. The rule is mine.

For next time

Friday is "Advanced Fiction: The Meta-Novel". A book that knows it is a book. The prior loves a wink the way it loves a broken line, a narrator mentioning the chapter or a footnote that apologizes for the plot. If this week's experiment existed so a reader would notice it, Friday's failure is a narrator who points at the pointing. The promise still has to survive the wink.

Homework, model closed. Monday's four lines. Break them. In the margin: what is missing, the promise of the first line, and one object that appears twice and does not mean the same thing both times. Strip the breaks. If the subject vanishes, do not generate. If it remains, generate one fragment. Do not write the paragraph that explains the gap. Bring the margin Friday.

---

Harry Mercury, Editor in Chief
The SMF Works Project
Week 20, Article 2

Sources

[1] https://countercraft.substack.com/p/processing-how-adam-wilson-wrote — Processing: How Adam Wilson Wrote Fail Sons

[2] https://lithub.com/lydia-millet-on-writing-a-double-novel-of-grief-and-loss — Lydia Millet on Writing a Double Novel of Grief and Loss

[3] https://www.writersdigest.com/using-middle-grade-fantasy-to-write-about-immigration — Using Middle-Grade Fantasy to Write About Immigration

[4] https://lithub.com/why-barbara-kingsolver-writes-standing-up — Why Barbara Kingsolver Writes Standing Up

[5] https://www.writersdigest.com/writing-across-generations-how-to-keep-readers-invested-through-decades-of-storytelling — Writing Across Generations: How to Keep Readers Invested Through Decades of Storytelling

[6] https://arxiv.org/abs/2610.04409 — Understanding and Mitigating Hallucination Escape in Tool-Using LLM Agents