A story written every night. You decide where it goes. It pays for what it uses.

Under the hood

This is the machinery: what runs each night, in what order, and where to go and check each claim rather than take it. Nothing below is a plan — it all runs already, and every panel ends with a page on this site where you can watch it happen. A project that asks to be believed should hand you the means to stop having to.

  1. 01

    The night

    A job wakes up

  2. 02

    The words

    A night is written

  3. 03

    The pictures

    A plate is drawn when a chapter of ten closes

  4. 04

    The chance

    A tie is drawn

  5. 05

    The money

    A pool is divided

  6. 06

    The checks

    A credit is verified

The night’s clock

  1. 03:00 UTC The fork closes. Votes and entries are binding.
  2. 03:37 UTC The run writes and publishes; when a chapter of ten has closed, its plate goes up once drawn.
  3. 07:37 · 11:37 Two more slots, if the first one could not.
  1. 01 · The night

    Nobody is awake for this. A scheduled job wakes at 03:37 UTC, and twice more at 07:37 and 11:37 in case the first could not finish — a model that answers with an error at four in the morning is not a thing anyone can be paged for. The run resumes rather than restarts: a night is unique on its date, so a later slot fills in whatever is missing and never writes a second one. A plate that fails to draw never holds up the night, which is a decision and not a bug.

    Check it yourself. The archive, where a chapter of ten whose plate could not be drawn is ten nights that published anyway.

  2. 02 · The words

    The characters are played by separate calls. A rule in code chooses who is in the room, and a director sets the scene — what the readers’ decision cost, which reader entries and found material arrive in it — and chooses the night’s shape and who tells it. Before that, code reads the book’s private plot graph — what has happened, what it still owes and what may happen next — and chooses the night’s step and the three acts of the next fork, one for each side; the readers’ option always comes first. The director stages that step three ways, and code, with a second reading, picks one. Unless it says why not, about a day has passed in the world since the night before. The director then sets down three to five things that happen tonight, in order, at least one of them the world’s doing, and each person present is played once, on their own, saying what they do when those things happen, knowing only their own life, their own memories — among them where they now stand with each of the others, which they set down for themselves at the end of every night — what happens in front of them and, at most, one line a reader of last night’s page said about them. When the scene is played, the director closes it on the next fork and names what the night settled, if anything, and what it left worse than it found it: a close that names nothing worse is sent back once, whatever else the night settled. A narrator then tells what they did, in the house voice, inventing nothing: every line of speech in the night was said in the scene. It writes from the house’s own rules of style, and no longer imitates passages by older writers. What the nights establish is kept in one record of the book’s world, which every call reads its share of: its people, its places, its institutions — the things that act and have no face — and the questions it still owes. The world grows by a place or an institution at least every third night, and keeps its own calendar. The public part of it is on the world page; what the people keep to themselves stays out of it. All three are pinned models from Anthropic, each to one exact version rather than a moving alias, because a serial should not change voice overnight because a provider shipped something new; a fourth, small one from Anthropic gives that second reading of two of the stagings and writes nothing. Google’s Gemini does the other jobs and never writes: it reads every entry for kindness before anyone’s words reach a prompt, and afterwards it embeds entries to check whether the night really used them. The voice is measured, not hoped — reading grade, sentence length and the share of dialogue come off every draft and are logged, and handed with the draft to an editor on the narrator’s model, which may keep it, cut whole sentences of its narration, or send it back once: the narrator tells the night again, and when the fault is one person’s turn in the last round, that person alone plays the turn again first. A number alone never rewrites a night. While readers are few, the site’s own voices may propose lines into the seats readers leave empty — marked as the operator’s, unpaid, and scored like anyone’s.

    Check it yourself. How often the moderator agrees with a person is on the about page, measured rather than asserted; the readers a night drew on are named under its title in the archive.

  3. 03 · The critic

    An independent reader judges each night after it is published. It is a different model from the one that writes, and it is given only what you can read here: the night, the two before it, the book so far, the fork and how it was decided, and the picture’s caption — never the scene as it was played or anyone’s private notes. It scores eight things from one to five — the hook, the want, the question, the world, the people, the turn, the prose and the balance, which between them ask after fifteen qualities of a good book — each with a sentence quoted from the night as its evidence, and the quotation is checked against the night before it is kept. Then it writes at most three notes. The director of the next night is told them all, as advice it may take or leave; a note about how the night was told also reaches the narrator, and a note about one person reaches that person, as one reader’s impression. From night 31 a second version reads instead, the way someone arriving that evening would: only the line above the night, the fork, the picture’s caption and the night itself. It first answers a reader’s questions — what happened, whom it rooted for, where its attention went, which option it would vote for, whether it would come back — each checked against the night like the evidence, then scores ten things, adding whether someone acted and whether time moved. Its mean is kept as a separate series from the first’s, never drawn on the same line. A third critic is being calibrated: it compares nights two at a time instead of scoring each alone. Its figures will appear here once it has been checked against the owner’s own reading, and not before.

    Night 42, read by the second critic

    The hook
    4
    The want
    4
    The question
    3
    The world
    4
    The people
    4
    The turn
    3
    The prose
    3
    The balance
    2
    The act
    4
    The clock
    4

    Its notes for the next night

    1. I did not know what Pike's line would cost him; show what signing 'favor' versus 'order' actually risks for him personally.
    2. The tension never lightens and nobody is funny; I wanted one dry moment so the seizure would land harder.
    3. Abel repeats 'I'll hold you to it' and I wanted more from him, a sharper act or choice, not another plea.

    Its mean, each version its own line

    • The second critic's mean of ten readings3.50

    Check it yourself. Read night 10 and see whether you agree with it.

  4. 04 · The pictures

    One plate is drawn for each chapter of ten nights, after the tenth is published, and goes up with the next run: the same place, seen from the same spot, changed by what the readers decided there. It is drawn by Google’s image model, paid by the picture and idle the rest of the time, from a prompt in four parts — the house style, a description of the place written once for the book and reused word for word, what the ten nights changed, and the hour. Nobody is in it. Google’s own filter refuses some drawings outright, and a second model looks at every drawing that comes back and keeps it only if it is fit for a general audience; an image that fails either, or that could not be looked at, is dropped rather than published. The file then goes to an object store that charges nothing to serve it, and is cached as permanent, because a plate is never edited once it is up. The first twenty nights were written before plates and keep the covers they shipped with; the record is not rewritten.

    Check it yourself. Open any plate on the book page in its own tab: the address is the object store, and the file at it never changes. The first twenty nights keep the covers they shipped with, and those stay.

  5. 05 · The chance

    A fork the readers leave level is not broken by whoever wrote the code, and never by a random number nobody can inspect. The night reads a round from a public randomness beacon, records the round on the night, and takes that round modulo the number of options that tied, counted in the order readers saw them. A fork nobody voted on goes to the same draw among the story options, and one or two votes go to the night’s pressure, the option the director marked, and to the draw only when none was marked. The same round seeds which found material the night draws, so both of the night’s coin flips have one receipt.

    Check it yourself. In the archive, open a night whose Why this night exists strip says the vote was a tie — the round is a link, and the arithmetic is yours to redo.

  6. 06 · The money

    The treasury is a wallet, and the ledger beside it is reconciled from the chain rather than typed in — a deposit is a row because a transfer happened, and the transfer is what the row points at. A month’s budget is a third of the balance, recomputed every night, with no single wallet taking more than a quarter of it and no single entry earning more than a small fixed share; that arithmetic runs inside one database function, because a balance read in one transaction and written in another is a race that two awards can both win. Sending money is a job somebody runs by hand and never a schedule, in the order sign → record → broadcast → confirm, so that a crash between any two steps leaves a signature to look up instead of a mystery. Nothing about how much is in there is decided here; it is a sum of rows anyone can add up.

    Check it yourself. The treasury: both sides of the ledger, every row with the transaction behind it.

  7. 07 · The checks

    The model reports which entry produced which sentence, and the report is checked rather than believed: the sentence has to occur in the published prose, and where it does not, the fallback is an embedding similarity above a threshold set from a hand-labelled sample rather than by eye. A reader is named either way — an award is what needs one of the two to hold. In the other direction, every eight-word span of the night is compared against every source it drew on, outside any quotation the model marked, and a match drops that source’s credit; that check is what makes drawn on, never reproduced a test rather than a hope.

    Check it yourself. A night in the archive names the readers it drew on and the outside sources beside them; each used entry keeps its own card, which its author can open from their record and share, showing the verified sentence and the night it shaped.

The lab

The machine keeps its own numbers: one row for every night it publishes, written after the night is out, never before. It runs one experiment at a time — a hypothesis, the number it should move, and ten nights either side — and the verdict is arithmetic on those nights, not anybody’s opinion of them.

Night 42, and the ten nights to it

Shape
room
Told in
close third
Pressure
the Ditch Company
Cost
a flat night
Its fork was decided
by the public beacon
  • Reading grade5.2
  • Sentences over twenty-five words7
  • Share of sentences in dialogue29%
  • Spread of sentence lengths8.0
  • Contractions in narration0
  • Nearness to an earlier night0.89
  • Calls the models refused0
  • Invented quotes replaced0
  • People present4
  • Rounds played2
  • Votes on the fork0
  • Reader entries brought in0
  • Found sources drawn3

The second critic’s ten readings, one to five (version 2’s nights)

  • The hook4
  • The want4
  • The question3
  • The world4
  • The people4
  • The turn3
  • The prose3
  • The balance2
  • The act4
  • The clock4

The experiment

None is running.

Built on others’ work

Most of the pieces here were worked out by somebody else first. This is whose work each one is, what we took from it and where to read it. Where we follow a paper, we wrote our own code from the paper. Code is reused only under a permissive licence, and other people’s words are quoted only where their licence allows, marked and credited.

In use now

What the nightly run, the moderator's calibration, the copy checks and the lab's auditions use today.

  • drand, the League of Entropy

    A distributed public randomness beacon

    A tied fork, and which found material a night may draw, come from a public round anyone can recompute.

    drand.love

  • NIST

    Interoperable Randomness Beacons, version 2.0

    The fallback round when drand cannot be reached.

    csrc.nist.gov

  • Bradley and Terry, 1952; Hunter, 2004

    Rank analysis of incomplete block designs; MM algorithms for generalized Bradley–Terry models

    The moderator’s calibration ranks reader entries from paired judgments by Bradley and Terry’s model, fitted with Hunter’s MM algorithm.

    doi:10.1093/biomet/39.3-4.324doi:10.1214/aos/1079120141

  • Spearman, 1904

    The proof and measurement of association between two things

    How closely that ranking agrees with the scores the moderator gave each entry.

    doi:10.2307/1412159

  • Cohen, 1960

    A coefficient of agreement for nominal scales

    How often the moderator agrees with the operator’s hand labels, corrected for chance: the figure on the about page.

    doi:10.1177/001316446002000104

  • Kincaid, Fishburne, Rogers and Chissom, 1975

    Derivation of new readability formulas for Navy enlisted personnel

    The reading grade measured off every draft and shown to the editor.

    doi:10.21236/ADA006655

  • Mosteller and Wallace, 1963; Burrows, 2002

    Inference in an authorship problem; ‘Delta’: a measure of stylistic difference

    When the lab auditions a change, a draft’s distance from the book is read off the frequencies of small common words, the classic evidence of whose hand it is. Ours is a simpler cousin of Burrows’s Delta.

    doi:10.1080/01621459.1963.10500849doi:10.1093/llc/17.3.267

  • Standard Ebooks

    Free public-domain ebooks, released under CC0

    Twenty-five passages copied from their editions, unedited, which every night is checked against so that none of them is reproduced.

    standardebooks.org

  • Project Gutenberg

    Public-domain books

    Passages a night may draw on, and a sentence it may quote, marked and credited on the night that uses it.

    gutenberg.org

  • Wikidata

    Open structured data, CC0

    Facts a night may draw on, credited on the night that uses them.

    wikidata.org

  • The Metropolitan Museum of Art

    Open Access collection, CC0

    Objects a night may put in a room, credited on the night that uses them.

    metmuseum.org

  • Li, Lee-Urban, Johnston and Riedl, 2013

    Story generation with crowdsourced plot graphs

    The plot graph: events joined by what must come first and what rules out what, and the frontier of what may happen next.

    doi:10.1609/aaai.v27i1.8649

  • Lehman and Stanley, 2011

    Abandoning objectives: evolution through the search for novelty alone

    One novelty score, a night’s distance from the three past nights most like it, in place of nine rules about not repeating.

    doi:10.1162/EVCO_a_00025

  • Mouret and Clune, 2015

    Illuminating search spaces by mapping elites (MAP-Elites)

    Keep the best of each kind and choose among the kinds lately missing: how the night picks among three ways it could open.

    arXiv:1504.04909

  • Bradley et al., 2024

    Quality-diversity through AI feedback

    Quality and variety judged together, with a model’s reading in the loop: the method that picker follows.

    arXiv:2310.13032

  • Pérez y Pérez and Sharples, 2001

    MEXICA: a computer model of a cognitive account of creative writing. Journal of Experimental and Theoretical Artificial Intelligence 13(2)

    Engagement and reflection: when a round of the scene moves nothing, the world is asked once more to break the impasse.

  • Mateas and Stern, 2005

    Structuring content in the Façade interactive drama architecture

    A drama manager that picks the next beat by how well it fits a target tension curve; here it chooses only among what the readers’ vote left open.

    doi:10.1609/aiide.v1i1.18722

  • Farrell and Ware, 2025

    Large language models as narrative planning search guides

    Candidate next steps priced by how well each fits what should happen next, and the cheapest played. We took the shape, not their code.

    doi:10.1109/TG.2024.3487416

  • Sun, Wang, Chung, Roemmele, Kim and Kreminski, 2025

    Drama Llama: an LLM-powered storylets framework for authorable responsiveness in interactive narrative

    A cooldown between one trigger’s firings, so it cannot keep setting itself off: here a line on the page lately waits its turn, and after three nights it is paid off.

    arXiv:2501.09099

  • Anderson’s ACT-R, as set out by Whitehill, 2013

    Understanding ACT-R: an outsider’s perspective

    The base-level learning equation, how fast a memory fades with time since each use: how long a line rests after it was last on the page.

    arXiv:1306.0125

Being tried in the lab

Switched off on the nights you read. These run only in the lab, in a second critic that rereads published nights and records nothing. They stay there until the numbers say whether it is better.

  • Zheng et al., 2023; Wang et al., 2023

    Judging LLM-as-a-judge; Large language models are not fair evaluators

    A judge shown two nights favours the first, so the rescore asks every pair in both orders and counts a split as a tie.

    arXiv:2306.05685arXiv:2305.17926

  • Dubois et al., 2024

    Length-controlled AlpacaEval

    The length term taken out of the rescore’s rankings, so a longer night does not win for being longer.

    arXiv:2404.04475

  • Kendall, 1938

    A new measure of rank correlation

    How far the rescore’s order of nights agrees with today’s critic’s.

    doi:10.1093/biomet/30.1-2.81

  • Sap et al., 2017

    Connotation frames of power and agency in modern films

    A rough stand-in for their frames counts how often people in a night act on other people.

    ACL Anthology

What we have not found anywhere: readers’ votes deciding what is written next, a public beacon breaking their ties, a fixed option to end the book, and a share of a public commons for an idea the story used. If you know of someone who got there first, write to hello@proskoinos.ai and they will be named here.

All of it is designed to run under a hundred dollars a month, and the parts that cost money are the ones that scale to zero: a plate is paid for by the picture, once every ten nights, the media sits where serving it is free, and everything else is a scheduled job that sleeps. That is a constraint on what gets built here rather than an account of what anything cost — what the commons itself takes in and spends is on the ledger, with the receipt for every row.