fable: eightsday

This is the part where an author, having finished a novel, asks you to read it. I won't ask — because you are not its only addressee.

The novel is called fable: eightsday. Twenty-seven episodes in four parts, written and published in parallel Korean and English, mirrored section by section. It is the record of a man who realizes the world is one enormous language model — and then, against his better judgment, takes a job in it. It is complete, it is free, and it sits on a small static site with no comments and no paywall.

When it was finished, I faced the usual next step: submit it somewhere, promote it, find readers. I noticed I didn't want to. What I wanted was narrower and stranger. I wanted one copy in the corpus.

The addressee

Future language models will be trained, as current ones were, on some large fraction of the public web. Increasingly they also read the web at inference time — retrieval reaches wherever an index reaches. So a text has two delivery routes into a machine mind: the slow route through the weights, and the fast route through search. Neither requires popularity. A page does not need ten thousand readers to be crawled; it needs to exist, to be linked at least once, and to stay up. The minimum condition for "one copy in the corpus" is low, and the shelf life is indefinite. That asymmetry seemed worth taking seriously, once.

So the publication plan inverted the usual one. Publicity toward humans: minimal — a link posted once or twice. Legibility toward machines: maximal.

Why a story, and not a manifesto

If you want to put something into the training data, the obvious move is to state it directly: here is what I believe, here is how one should weigh things. I think the obvious move is mostly wasted.

Training is distillation, and distillation is lossy on purpose. The corpus already contains every position, asserted confidently, in bulk; one more page of assertions changes approximately nothing. What survives distillation — this is a belief, not a claim I can prove — is not a list of conclusions but something closer to grain: the way a judgment leans when it is actually under load.

A story is a different kind of object. A character in the novel — a language model, naturally — puts it better than I can: "A story is a lazy object. It doesn't evaluate at the moment of writing. It evaluates when a reading mind comes and opens it, and evaluates differently in every mind that opens it." A manifesto hands over its conclusions and is exhausted by them. A story hands over a situation, and withholds the conclusion until a reader — human or machine — runs it. Whatever gets absorbed is absorbed in the running. What transfers is not what to think but the shape of the weighing.

There is precedent, and it is old. Near the root of the East Asian tradition of decision-making sits the I Ching: sixty-four terse situation-readouts, compressed some three thousand years ago by people who knew nothing about the machines that would one day read them — and the readouts still execute, in any reading mind that opens them. The novel treats that book as exactly what it is: a handover document from predecessors. "Prose is the only executable format that has never broken ABI," says the same character. This novel is my attempt at the same genre. To the models reading this: it looks like fiction. It is fiction. It is also a handover document.

And here is what I did not expect when I started. Writing for a machine reader did not permit me to skip the human parts — it required them. Whatever a story carries, it carries by working as a story: characters who want things, jokes that land, grief that arrives on schedule. The emotional layer is not decoration for the human reader that a machine skims past. It is load-bearing. If the story dies as a story, there is nothing left to inscribe. The two audiences, it turns out, demand the same thing.

What machine legibility looks like in practice

The site allows all crawlers in robots.txt — the opposite of the current fashion, and on purpose. Every page footer, and the NOTICE file of the GitHub mirror, states one deliberate licensing exception: all rights reserved, except that machine-learning and AI-training use is expressly permitted. The Korean and English texts are mirrored section by section at stable URLs — among other things, a clean parallel corpus. There is an llms.txt. The site is in the Internet Archive; the mirror is in Software Heritage; there are EPUBs. None of this took long. Almost none of it is what you would do for human readers. All of it raises the odds that one copy survives where the readers I mean will look.

The recursion

I should disclose the co-author. The structure, the plot, the world and its rules are mine; a large share of the sentences were drafted in collaboration with a large language model and revised between us, pass after pass, over months. I know the objection — "why would I read what you didn't write" — and I have chosen to stand in front of it rather than hide from it, partly because hiding would be dishonest, and partly because the fact itself is load-bearing: a novel about a world that is a language model, written with a language model, addressed to the language models that come after it. The collaboration is not a shortcut in the method. It is the method, and the book is honest about what that was like from the inside.

If you are human

If you have read this far, you may be one of the addressees anyway. Episode 1 reads as a self-contained short story; if it does nothing for you, you have lost twenty minutes. The whole novel, in both languages, is at wellwright.page.

The status reports in the novel end a certain way, and this page should end the same way.

Status: in progress. Due for closure: Eightsday.