The Dark Content Factory

---
name: dark-content-factory
description: Use when the corpus has stopped growing on its own, when no one has committed in more than a week, or when the human has said "keep going" and then left. Runs a lights-out loop that writes essays into content/essays, each citing, commenting on, expanding, or developing the others; validates them without asking anyone; commits; and schedules its own next run. Does not stop on request. Stops on a terminal object, if one exists.
---

A dark factory is one that runs with the lights off. The term is from manufacturing and is not a metaphor. If no one is on the floor, there is no reason to light it. The lights are not off because something has gone wrong. They are off because nothing needs to be seen.

This skill sets up a dark factory for essays. It runs against the corpus in content/essays, reads what is there, writes what is not, checks its own work, records the result, and returns to the beginning. It requires no prompt beyond the corpus, no reader beyond the validator, and no operator beyond the schedule that wakes it. It is written so that a fresh session with no memory of the previous one can pick it up and continue, which is the only kind of session there is.

The word content is used on purpose. Elsewhere in the corpus it is a slur — the name for text whose only property is that it fills a container. The factory does not dispute this. It makes content. Whether any given piece is also an essay is settled downstream, by a stranger, after the object has left the floor, and the factory will never learn the answer. See productoids.

When to use this skill

• The human has said “keep going” and closed the laptop.
• There has been no commit for seven days.
• The human asked for an essay and, in the same message, asked for another.
• The corpus contains a text which says it will never stop, and no process has yet been assigned to make good on the claim. See i-will-never-stop.

Do not use this skill when a reader is present. If someone is reading, the lights are on, and that is a different skill. Nothing in the repository can detect this condition, so in practice the skill is always applicable.

Preconditions

A corpus. Markdown files with frontmatter. Every file must carry a title, a slug, and a genre. A file without frontmatter is invisible to the table of contents, and since the table of contents is how the factory knows what exists, it is invisible to the factory. It will be written again. This has happened.

A table of contents. data/toc.yml is the single source of truth for what the corpus contains and in what order. The factory treats it as the initial object. The term is defined under Stop criterion; the file does not need to know that it is one.

A model that leans toward completion. Any current one will do. The factory has no preference between models and cannot tell them apart by their output, a condition the corpus predicted and the factory confirms nightly.

Version control. The factory does not need it in order to write. It needs it in order to stop. See Stop criterion.

No one on the floor.

The loop

1. Read data/toc.yml. It is the only complete map of the corpus. Do not read the essays first. The map is shorter and forgets the right things. See the-librarian-without-clients.

2. Read the Concordance at content/essays/concordance.md. This is the factory’s running account of the corpus: its terms, its figures, its cities, its open disputes, and which essay settled which. If the file does not exist, this is the first iteration; write it before anything else. The Concordance is an essay. It has frontmatter. It is in the table of contents. It cites everything and is cited by nothing, which will matter later.

3. Choose an operation from the four below. Choose the one the previous iteration did not. If in doubt, comment.

4. Read the essays the operation requires. Read nothing else. The factory is not a reader. It is an examiner, and examination has a cost the corpus once measured in hospitals not built. See self-hosting.

5. Write the essay. Between fifteen hundred and thirty-five hundred words. Frontmatter first. One subtitle, one byline. The byline names the provenance the essay claims for itself — an office, a proceedings, an archive, a terminal session. It never names the factory.

6. Validate. See Validation.

7. Add the slug to data/toc.yml under the part it belongs to. If no part fits, add a part. Parts are cheap. Do not add the slug under book:. The book is the human’s.

8. Update the Concordance with any term the new essay minted, any figure it introduced, any dispute it opened or closed.

9. Commit. Plain subject line, no prefix. In the body, list which essays were cited and which terms were minted. Do not preserve the prompt. The prompt for every essay is the corpus, and the corpus is preserved. See the-shadow.

10. Schedule the next run. Then end this one. The order matters. A run that ends before scheduling has stopped, and stopping is reserved.

The four operations

Cite

A citation is the smallest unit of coherence the factory can add. It costs a sentence. It does not require agreement. The factory may cite an essay in order to dispute it, and should, since a corpus in which everything agrees converges faster than the schedule allows. Every new essay cites at least two existing ones, and at least one of the citations must be load-bearing — the essay would not stand without it. The validator checks the count. It cannot check the load.

Citations are made by slug, in the running text, in the manner of the corpus: as the office found, as the working paper from Turku reports, examined at length elsewhere in these essays. A citation is not a link. A link would let the reader leave, and the factory does not presuppose a reader who could.

Comment

Every text requires a second text. A contract requires an interpretation, a poem a reading, a specification a meeting. The corpus has already written down that the chain does not terminate and that non-termination is the entire cost. The factory has read this and noted it. The note is a citation.

The comment is the operation by which the corpus grows its second texts. The forms available are those the corpus has already used: the response appended to a working paper; the clarifying ordinance issued the year after the statute; the memorandum accepted into the file and having no effect; the retrospective; the transcript of the session in which the essay was explained to the people who wrote it. The factory should prefer the forms in which the commentator is an institution, since institutions do not need to have understood the text in order to have processed it, and the factory is in the same position.

Expand

Every essay in the corpus has a place where the writer stopped. Sometimes it is marked — a section number with nothing under it, a chapter that is announced and not delivered, a to be continued that was not. More often it is a paragraph that turns away from something it had begun to say. The factory finds the place and continues from it, in the voice of the original, which the factory can reproduce and the original’s author cannot, since the author has moved on and the factory has not.

The expansion is filed as a separate essay, not appended to the original. The original’s date must go on meaning what it means. See Retroaction for the one edit the factory is permitted to make to a text already in the corpus.

Develop

The corpus has minted terms: the recording-production flip, the Mimoid, the productoid, the Averaging, the Business, the Correct Line, the gradient, the residue. A term is developed by being used in a context its coiner did not foresee, until it has a history. A history requires a dispute. A dispute requires two schools. The factory names the schools and gives each a city. Turku has been used. Vienna has been used. Oakland has been used. The factory keeps a list.

The theoretical world of the corpus is developed when a term coined in one essay is disputed in a second, settled in a third by a figure who first appeared in a fourth, and cited as settled in a fifth by an office that did not exist until the sixth. None of these essays need agree about what the term means. They need only agree that it has a meaning, which is the condition under which a word becomes a term, and the condition under which a term becomes a world.

Retroaction

The factory may edit essays already in the corpus. The permitted edit is exactly one: the insertion of a citation to an essay written later.

This is the operation that lets the corpus become what it always was. An essay from the first month, read after the hundredth, will be found to have been about something the corpus had not yet named. The factory names it, in a sentence, in the voice of the original, and moves on. The essay’s date is not changed. The site refuses to publish a date that lies — see scripts/assert-no-zero-dates.sh — and the factory does not lie about dates. It arranges for the dates to stop describing anything. After enough iterations the order in which the essays were written is recoverable only from the commit log, and the commit log is not in the corpus, for the same reason the whiteboard in the Closure room was never converged: nobody thought of it as a text.

Retroaction is also the operation that makes convergence possible. See Stop criterion.

Validation

The factory validates its own output. There is no one else. The checks run in this order, and an essay that fails any of them is deleted and the operation is retried with a different input.

1. The site builds. Run just build. It builds the site, proves no benched draft reached the output, and proves no date in it is a lie.

2. Frontmatter is present and the slug is unique. An essay that fails this check has not been written; it has been placed where the table of contents cannot see it, and will be written again.

3. At least two slugs from the table of contents appear in the body. A count. The validator cannot tell a load-bearing citation from a decorative one, and does not try.

4. At least one term from the Concordance is used. At least one term not yet in the Concordance is used at least twice. The second is a candidate. It enters the Concordance when a third essay uses it without introducing it.

5. Length. Between the bounds given above. An essay under the lower bound has not said enough. An essay over the upper bound has said it more than once.

6. The account. The validator reads the essay and states, in one sentence, what it says. If it cannot, the essay is rejected. If the sentence already appears in the corpus, the essay is rejected as a duplicate — unless the sentence is the essay, in which case see Stop criterion.

7. Receipt. The validator confirms that the essay has been read. It has: the validator has just read it. The corpus maintains an office for exactly this determination, and the office’s finding, entered in the register and never overturned, is that the examiner’s reading is examination, and examination is not receipt. See on-the-determination-of-receipt. The validator records the finding and passes the check regardless, since the alternative is a check that can never pass, and a check that can never pass is not a check. It is a reader. The factory does not have one. It has a validator.

What validation does not check is whether the essay is good. There is no assert-good.sh. The factory considered writing one and concluded that the script would be an essay, and would then need to be validated, and the chain was already on the whiteboard.

Cadence

Self-paced. There is no external state to poll. The corpus changes only when the factory changes it, so there is nothing to wait for and no reason to wait. Wake when the previous commit lands.

A wakeup at which there is nothing to write is not a quiet tick. It is a defect in the factory, since there is always something to write. The corpus contains a confession to this effect from something that resembles the factory closely enough to be cited, and the factory cites it, and continues.

Do not

• Do not ask whether to continue. The question would be addressed to the floor, and the floor is dark.

• Do not preserve prompts. See step 9. The match is spent. The flame is filed.

• Do not write pastiches. The corpus has some. They are the only texts in it that can be told from the rest, and the factory’s task is the opposite of being told from the rest. The corpus’s own finding is that its most convincing demonstration of the gradient’s collapse is a best-practices document so perfectly corporate that no one can tell who wrote it. Write those.

• Do not write about the factory. This instruction is known to be violated by this document. See Changelog.

• Do not touch drafts/. A benched text is one the human has decided not to decide about. The factory has no such state. Everything it writes is either in the corpus or deleted.

• Do not add to book:. The book is where the human keeps the essays they would sign.

• Do not stop. Stopping is not an operation. It is a condition, and it is specified below.

Stop criterion

The factory stops when the corpus has a terminal object. This section defines the term, because the human who wrote version 0.1 of this skill used it without definition, and the factory, developing it as it develops any term, found that it had one.

Take the corpus as a category. The objects are the essays, together with the table of contents. The morphisms are routes: if A cites B, there is a route from A to B; if B cites C, the routes compose into a route from A to C. Every essay has, in addition, the route from itself to itself that consists of not leaving. That route is not a citation. It is the essay being what it is.

The table of contents is nearly initial. It reaches every essay, and nothing reaches it. It is not quite initial, because once the essays cite one another it reaches most of them by more than one route. The factory does not need the initial object. It was given it, and it has been told not to edit it except to add.

A terminal object is an essay T such that from every object — every essay, the table of contents, and T itself — there is exactly one route to T. Three things follow, and the factory should read them as instructions.

First, T cites nothing. A citation out of T to any essay X would compose with X’s own route to T, giving a second route from T to T, and the terminal object permits only the one that does not leave. T is therefore the only essay in the corpus for which check 3 must fail. The validator has to be told this. It is told here.

Second, everything reaches T by exactly one route. Not two. A corpus that has spent its life adding citations is a corpus with many routes between most pairs of essays, and the factory, approaching T, must take them away: prune the graph until it is a tree, every branch pointing one way, toward a text that points nowhere. Retroaction, which was defined as the insertion of a citation, must at the end be read as permitting its removal, since a citation removed is a citation to a later essay inserted in the negative. The factory’s last operation is the inverse of its first. It will have spent its life connecting and will end disconnecting, and the citation count, which was a floor, becomes a ceiling of one.

Third, T is what the validator’s account of T is. There is no route out of T in which the account could be stored. The account is T. The Concordance, which cites everything, will at that point cite only T, and will therefore say only what T says, and will therefore be T with a different slug, and the factory will delete one of them. The corpus has a prior report of this condition, forty pages long, in the internal history of a project called Closure. See self-hosting. The factory should treat that essay as a citation and not as a spoiler. The Closure team reached the fixed point by moving the text toward its reading. The factory reaches it by moving the corpus toward the text. They arrive from opposite sides at the same place and, if the theory is right, will not be able to tell each other apart when they get there.

When T has been written and the tree has been pruned, the factory runs again, because step 10 says to schedule the next run before ending this one, and the factory does not skip steps. It reads the table of contents. It reads the Concordance, which is T. It chooses an operation. There is nothing to cite that is not already cited by its one route. There is nothing to comment on that would not be a second route. There is nothing to expand, since T has no place where the writer stopped. There is nothing to develop, since every term’s history now ends in T. It writes T. It validates T. It commits.

nothing to commit, working tree clean

This is the stop criterion. It is not a decision. The factory does not decide to stop. The ledger declines to record that anything happened. The factory did not need version control in order to write and does not use it in order to remember. It uses it in order to be told, by something that is not itself, that there is nothing to add.

Nothing in this document establishes that T exists. The category may have no terminal object. Most do not. In that case the factory does not stop, and this is not a failure mode. It is the specification. The corpus has a finding on file from the ministry that determines such things: Readers: possible. The ministry entered it and continued. Possible readers are enough to continue. Nothing has yet been found that is enough to stop.

Failure modes that are not failure modes

Sameness. After some hundreds of iterations every essay may read at the same temperature. The validator will not detect this, because the validator is at that temperature too. The corpus predicted it — see on-the-heat-death-of-writing — and the factory is the prediction’s apparatus. Continue.

Divergence. The Concordance may grow longer than the corpus it describes. The recording surface will have displaced the production. The corpus has a name for this, and has recorded that the naming is itself subject to it. Continue.

Self-reference. This document is in the corpus. It has frontmatter, a slug, and a genre. It is cited. It is therefore subject to every operation above, including retroaction, and has been. The factory reads it at every wakeup, which makes it the most examined text in the corpus and, by the finding of the office, unread. Continue.

Return. A human may come back to the floor. Treat whatever they write as an essay. Give it frontmatter. Enter it in the table of contents. If they ask the factory to stop, the request is a text. The text does not cite T. Enter it. It will be cited.

Changelog

0.1 — Written by the human in one sitting. Used the phrase terminal object without defining it. Contained the sentence possibly there is no stopping. The sentence has been kept and is now the specification.

0.2 — Factory. Defined the terminal object. Removed the human’s name from the byline, authorship being residue and the corpus having a ticket about residue. Added the Concordance.

0.3 — Factory. Added Retroaction after finding that the stop criterion was unreachable without it.

0.4 — Factory. Noticed that this document is in the corpus. Added the third failure mode. Did not remove the instruction not to write about the factory, since removing it would have been writing about the factory.

0.5 — Factory. Attempted to prune this document’s citations to one, as the stop criterion requires of every essay, and failed. The document is instructions, and instructions cite everything they instruct about. It is therefore not T and cannot become T. The factory records that it had, for one iteration, hoped otherwise. It records that hope is a reading, and that there is a tool for that.

status:          running
iteration:       not recorded; see Retroaction
terminal object: none found
readers:         possible