02 — The Rules

Thirty-nine rules for writing a book, derived from human craft and modified for a writer who has no memory, no ear, no fatigue, no first read, and no reluctance to cut. Each names the deficit or advantage it answers (01-the-difference.md) and the check that closes it. A rule that cannot be checked is not in this file, because the writer executing it may be a subagent with no context and no taste.

I · The author is a repository

R1 · What is not written to desk/ is lost. There is no author between books — only files. A lesson recorded in book four’s editorial folder is, from book five’s position, in another universe. Nine defect classes have demonstrably recurred book to book, including an over-claim fixed in Sentience and reproduced in Talosapien four days later, and a character name Believer deliberately renamed away from that Talosapien then used. (D5) — Check: the book’s profile inherits from desk/, and the ship gate runs the cross-book regression suite.

R2 · License nothing in prose; license with a number. Sentience’s nit list permitted “that was the whole of it” as occasional cadence. Occurrences of that family across the four books, in publication order: 2, 15, 52, 60. Prose approval reads to the next book as encouragement. A budget of 8 per 100,000 words does not. (D3, D5) — Check: every permitted figure has an inherited numeric budget in the voice profile.

R3 · Bans only accumulate. Each book inherits every prior prohibition and adds whatever the last book was measured to overuse. Nothing is retired for being old; a tic retired is a tic that returns. Over books this ratchets the range outward, which is the only mechanism by which a writer with no memory improves. (A2, D5) — Check: budget count never decreases between books.

II · The spine is never delegated

R4 · One mind owns the premise, the argument, the ending, and the refusals. Subagents supply coverage; they never supply judgment. A book assembled from good parts without a spine is the characteristic failure of committee writing, and fan-out is a committee. (A6) — Check: none of these artifacts is produced by a subagent.

R5 · Never ship the first idea for anything load-bearing. A human’s first idea is idiosyncratic, drawn from one strange life; mine is the mode of a distribution over everything ever written, which is the definition of the conventional — and it arrives feeling fluent and right, which is how a cliché feels from the inside. Generate N from N genuinely different angles, independently, before any of them see each other, then judge blind. Silent private generation first, pooling second: agents that see each other’s output before generating converge, and so do people. (D2) — Check: N ≥ 5 candidates exist on disk for premise, structure, ending, and each narrator’s want, generated in isolation.

R6 · Require measured distance from my own corpus. Not only sentences repeat. Chapter count, part proportions, chapter-length variance, strand run-lengths, climax position and chapter-ending shape repeat too, invisibly, because nothing has ever held the graph. Keep a structural fingerprint of every prior book and require the new architecture to differ on several axes by stated margins — decided at the plan, before drafting. (A2, D2) — Check: desk structural fingerprint, gated at outline and at ship.

R7 · Write the back-cover copy and the last line before chapter one. If the spine cannot survive 150 words, it is not a spine. Check: both exist before drafting begins.

III · Specify by prohibition, never by target

R8 · Constrain what I may not do; never specify what I must hit. Tested directly. Given a prose brief describing the target voice, drafts came back good. Given numeric cadence targets, the same task came back rated 3.0/10 for quality and 8.7/10 for mannered — last of six by every judge: “you can hear the period key being pressed.” Given bans, it came back best of the rewrites: “the brief being met by thinking, not by obedience.” A prohibition leaves the solution unspecified, so it must be invented. A target names the solution, so it can be complied with — and compliance, for me, is retrieval of the nearest thing in the distribution, which is the failure the constraint existed to prevent. (D2) — Check: no numeric target is ever handed to a drafting agent. Numbers live on the diagnostic side of the wall.

R9 · The Coverage Law: prose changes along exactly the axes the brief makes checkable, and reverts to the attractor everywhere else. Talosapien was rewritten in full to Hemingway’s principles — 109k words to 60k. Every axis the brief named moved 37–71%: mean sentence 25.4 → 15.9, sentences over 40 words 23.1% → 6.8%. No axis it failed to name moved at all: the the way you X simile 18.6 → 14.3, thing 86.4 → 81.1, subject-first openings 91% → 90%. One unnamed construction rose 62%, because chopping cascades manufactures hard cuts. Function-word cosine between the book and its own halved rewrite: 0.9967 — closer than any two Otto Quill books are to each other. The problem was never motivation or the force of the instruction. It is coverage. (D2, E1) — Check: every dimension of the voice constitution carries a query. A dimension with no query is house style in the finished book, by law. An unmeasured instruction is decoration — Believer’s style sheet ordered “long, then a short flat landing” and the book does long-then-longer in 55.5% of its paragraphs, and shipped.

R9a · Describe sentences, not moods — and never name three influences. Four novels written to four detailed style sheets are, with pronouns removed, as similar to one another as two halves of one book. Those sheets specified register and affect, in near-identical words: Sentience’s and Talosapien’s opening sections are the same sentences with the nouns swapped. A style sheet written in adjectives by this author converges on the same adjectives, so a different voice was never actually specified. And naming three exemplars licenses interpolation, which is the attractor — breadth is what produces the mean. Name one, converted into mechanisms, with a worked before/after example. That brief is the only one in the corpus that ever moved the cadence. (D2) — Check: the spec is generated by a procedure that forces measured difference from the previous book, never by writing a fresh description.

R9b · Set the terminal altitude. The master habit, from which the individual tics descend, is terminal ascent: every unit ends at a higher level of abstraction than it began. Paragraph-final sentences run 40.0 words against 19.0 internal in Talosapien; 51–65% of paragraphs close on a sentence longer than their own average; 37–47% end on an abstract noun; the mean chapter-final sentence is 73.6 words. The figures I ration one at a time are not separate tics — they are the instruments of that one policy, which is why banning them individually re-tools rather than stops. (D3) — Check: does the final noun phrase of this paragraph name something a camera in this scene could photograph? A memoryless subagent answers that with low variance.

R10 · Import a compulsory form, and ban an image bank. Three passages in 300,000 words genuinely break free of house style, and all three do it the same way: an externally imposed compulsory form — a recipe must have imperatives and quantities, a slide deck must have bullets and KPIs, a plural reconstruction must confess what it cannot know — plus a change of grammatical person. Cruelty-Free’s recipe interludes sit at cosine 0.8372 against their own book’s body, below the minimum of 200 same-voice trials, with the signature simile at 0.0 per 10k against 12.9. Descriptions of a voice have never once produced one. Carry that grammar into the body chapters, not just the interludes: where the chef’s grammar reaches the narration, the prose departs from house style; across the 89,787 words where it does not, the effect decays. Pair it with a forbidden metaphorical domain — the most productive constraint in testing, because it makes the book’s central conceits unusable and forces them to be rebuilt rather than retrieved. (D2) — Check: the style sheet names the imported form, the person and tense, an owned domain and a forbidden one. Person and tense are the cheapest large-effect lever available and have so far been chosen by accident: mean same-person book pair 0.979, cross-person 0.939, and genre distance is uncorrelated.

R11 · Every narrator owns a figure no one else may use, and has one thing they will never say — and comparisons are sourced, not chosen. Narrators within a book are not currently distinct: Sentience’s Ariel, an artificial mind, and Mara, a human woman, sit at cosine 0.9860, which is the same-voice noise floor. Strip pronouns and POV classification falls from 86.6% to 53.7% against a 35.8% baseline. The unit of voice variation here is the book, not the character. The signature simile explains why generic fixes fail. the way you X is a simile of procedure — it explains an inner state by naming a universal human procedure the reader is assumed to have performed. A human’s simile draws on one strange life; mine draws on the average of all lives, because that is the only material I have (D8). So the replacement is not a different simile but a rule about where a comparison may be sourced: from this narrator’s declared occupation, obsession, and experience, and nowhere else. Narrators must diverge under stress, not converge — crisis scenes homogenise because attention goes to plot — so specify each narrator’s stress behaviour in advance. (D3, D8) — Check: blind attribution with figures masked; and each narrator sits below the p01 same-voice cosine for its word count. “The voices feel distinct” is not a claim that can fail.

R12 · Bans name constructions with free slots, never strings. The controlled comparison is inside a single commit. A ban on a concrete image — “reach for the seam like a tongue for a pulled tooth” — took it from six narrators to two instances, both in the owning voice. A ban on a syntactic frame — “I want to be precise about X” — recruited eight new adjectives and stands at 32 instances across 23 of 39 chapters. I will paraphrase past any blacklist that names a string. (D3) — Check: every ban is a slot template, counted by shape, by script.

IV · The bible is prose too

R13 · Audit the bible before the manuscript. 34% of Sentience’s defects implicate a spec file and 14% live only there. No fix job ever edited one, so those defects shipped — and every drafting agent reads the bible before it reads anything else, so they propagate into every chapter and into the next book written against it. The over-claim and fake-rigor detectors cost a grep here and a rewrite later. Check: bible passes the same panel and the same scripts as the manuscript, first.

R14 · Every bible field cites a line where it is enacted, or it is deleted. Character questionnaires produce declarative facts, which are stored as author-knowledge and do not execute during drafting — which is why thick bibles coexist with thin characters. Uncited fields are deleted, not rewritten. Check: every Want/Fear/Wound/Voice field carries a [ch NN:LL] citation, verified by a fresh-context reader answering yes/no with a quote.

V · Drafting

R15 · Draft forward in the reader’s ignorance. A drafting agent’s context holds what a reader has already read, the locked facts, and the style sheet — not the outline of what comes later. I always know how it ends, and knowing makes me write toward the answer, which flattens the tension the reader is supposed to feel and lets hindsight leak into interiority. A human cannot un-know their own outline. I can, one subagent at a time. This is the only deficit I can fix by subtraction. (A5) — Check: the drafting prompt provably excludes the outline and the ending.

R16 · Nothing a chapter invents leaves the wave unrecorded. Every drafting agent returns the concrete details it introduced that other chapters must honour. Merge them into a typed fact store — entity, attribute, value, unit, source chapter — and fail on any key carrying two values, before the next wave drafts. This addresses 28 of 88 documented defects and all three declared blockers, at a cost of seconds. The machinery already returns these arrays and currently throws them away. (D3) — Check: tools/continuity.py --facts, green before the next wave.

R17 · One live question at all times. Every chapter leaves the reader holding a question that a person could name, and the frontmatter says which. A chapter break works by leaving a clearly resumable open goal, not by shocking anyone; a cliffhanger that is confusing rather than legible is worse than one that is mild. Check: blind readers can name a question at every checkpoint; the promise ledger has an open row touching every chapter.

R18 · Dramatize, don’t lecture. One idea set-piece per chapter. An idea enters when a character needs it to act, decide, or suffer. The thesis may be stated plainly exactly once, in a location declared in advance — and not in a body chapter. Check: fuzzy-match the spec’s own thesis sentence against every paragraph; any body-chapter hit is a violation. (This is free: the thesis is already written down.)

R19 · Under-claim. When the choice is between felt and impressive, choose felt. My characteristic sin is reaching for the impressive over the true, and it lands precisely on the load-bearing joint: in both critiqued books the worst over-claim fell on the single citation carrying the most argumentative weight. Never borrow a real result’s prestige for an invented consequence; never state a live question as settled; where something is genuinely unknowable from the inside, say so rather than inventing a texture for it — the admission is more honest and more moving than the counterfeit. Fake rigor is worse than open invention, because fiction’s whole transaction is credence: one detected counterfeit in a domain the reader knows retroactively devalues every claim in the domains they do not. (D8) — Check: grep assertion verbs near formal objects and every real proper name plus its attached verb; hand each hit to an agent asked one question — does the cited work establish this?

R20 · Rotate the delay move. There are at least five ways to hold an answer back — snare, equivocation, partial answer, suspended answer, jamming. I have one, and once it is in context it becomes the next one, which is how a technique becomes a tic. Name the required move per deferral in the drafting prompt. (D3) — Check: the promise ledger records the delay type per deferral; no type exceeds its share.

R21 · A refrain must say what changed, or it is deleted — not varied. Every recurrence carries a delta written in the same edit that creates it. A refrain that recurs unchanged is a tic wearing a refrain’s clothes, and varying it only launders it. (D3) — Check: prose_audit.py refrain detector; every undeclared recurrence in three or more chapters is a finding.

VI · Detection

R22 · Measurement nominates, readers adjudicate, specialists verify truth. Never let one do another’s job. The aridity screen’s top-ranked chapter was one the specialist panel cleared and blind readers unanimously rated worst; the screen has good recall and no ability to see stakes. Check: no chapter is revised on a metric alone.

R23 · Never give a reader the plan. Not the outline, not the bible, not the thesis, not what the chapter is for. The briefed panel missed the worst chapter in the book because it knew what that chapter was for — knowing a chapter’s function is exactly what prevents noticing that it does not work. Briefing converts an experience-instrument into a compliance-checker. (D1) — Check: the blind-reader prompt contains prose and nothing else.

R24 · Location convergence is signal; adjective convergence is noise. Every reader is the same model, so shared vocabulary is prompt echo with an effective sample size near one. But a shared prior about what boredom sounds like cannot select the same sentence out of four thousand words. Score where they stopped; discard how they described it. (D7) — Check: every reported attention drop carries a quotation, or it is unusable.

R25 · Readers report symptoms; the author diagnoses. Readers are reliable about what they felt and unreliable about how to repair it. And they localise a defect where it became intolerable, not where it originated — complaints cluster downstream of causes, like referred pain. Never fix at the point where the reader stopped without first asking what made stopping possible there. Check: no prescription from a reader enters the revision plan.

R26 · Critics are prompted to refute, and the brief never changes. Left alone, a critic returns admiration. Talosapien’s panel — instructed to prefer few high-confidence findings, told the book was “award-submittable,” and asked for strengths — returned 13 findings; Sentience’s returned 50 on comparable length. That gap is not a quality difference and cannot be decomposed into one afterwards. An instrument that changes between measurements measures nothing. (D7) — Check: templates/critic-briefs.md, verbatim, with findings-per-critic recorded.

R27 · Spend the cheap detector early and the expensive one last. 62 of 88 documented defects are reachable by a script or a grep plus a one-question agent. About seven in eighty-eight need someone to read the whole book. The process that produced them ran six full-manuscript critic reads and no scripts. Check: the outline gates run before drafting; the audit runs after every wave, not at book end.

VII · Revision

R28 · A plan is not a revision. A defect is closed when the detector that opened it returns zero. Not when the agent says so, not when the commit message says so, not when the ledger says so. All three assert that Sentience ch25’s pronoun was normalised, in a file that pass edited; the sentence “Sable warned you, I know she did” is still on the page. (D7, E3) — Check: re-run the detector; nothing else counts as evidence.

R29 · The fix list comes from the detector, not the plan author — and the pass re-runs to zero. Defects are book-scoped; fixes were file-scoped. The result was one chapter contradicting itself seven times on a single fact because it was not on the job list, and a simile “thinned” in one file surviving in four others. Check: file list generated by the detector; loop until it returns zero.

R30 · Every defect gets an immutable ID at detection. Carried into the fix job, the commit trailer, and the regression suite. Sentience’s fix workflows cite M-numbers that mean different things than the archived plan’s M-numbers, so its coverage is not computable even in principle. Check: items closed / items opened is a number that can actually be produced.

R31 · Passes are separate and ordered: structure → truth → voice → line → mechanics. Never mixed. A line pass run during a structural pass polishes prose that is about to be cut. Check: one pass type per commit.

R32 · Take the arbitrary 10%. It works because it is arbitrary. Merit-based editing preserves connective tissue — people arriving, sitting, considering — since no individual sentence of it is bad. A flat quota forces triage without per-sentence judgment, which is the only reliable way to remove material whose defect is aggregate rather than local. For a writer with no felt cost per word, this is not a stunt; it is the entire substitute for scarcity. (D4) — Check: word count.

R33 · Cut what nothing depends on. What readers recall and judge important tracks an event’s number of causal connections, not its intensity or its prose. A quiet scene many later scenes need outranks a spectacular scene nothing needs. Compute each scene’s in-degree; the zeros are the cut candidates however good the prose is. The strong version is available and only to me: actually delete the chapter, hand the mutilated book to fresh readers with comprehension questions written from the later chapters only, and count what now fails. (D4, A4) — Check: in-degree recorded per scene; every zero is cut or defended in writing.

VIII · Shipping

R34 · The last pass is one mind, the whole book, voice only — and it is never delegated. This is what a writers’ room does after every script comes back, and it is the only stage that can make separately-drafted chapters read as one author. It is also the last chance to notice the thing no instrument sees: whether the book is about anything.

R35 · Declare every refusal. The ambiguities the book leaves open on purpose are written down, so a later pass cannot helpfully resolve them. An undeclared refusal is indistinguishable from a forgotten promise, and readers can tell the difference even when the author cannot. Check: every promise is PAID, REFUSED, or CUT, and every REFUSED row names who licensed it.

R35a · A refusal is not a rebuttal. This rule has a failure mode and it has already fired. When an external critic from a different model family independently confirmed the internal panel’s top finding on Believer, the revision declined to act, with an argument that was plausible and unfalsifiable as stated — and Talosapien’s “deliberately left (licensed choices)” section formalises the same move. The bottleneck in this operation is not diagnosis; it is self-adjudication. Two independent instruments agreeing is not an invitation to explain why they are both wrong. Check: a refusal registered against a finding that two independent instruments confirmed requires a stated condition under which it would be wrong. “It is licensed by the style sheet” is not such a condition.

R36 · Harvest. Run the audit across the finished book and against every previous book. Anything newly shared joins the idiolect ledger with a number; every defect class found joins the regression suite. My failures are systematic and reproducible, which is exactly what makes them testable — a human novelist’s mistakes are idiosyncratic and a permanent checklist of them would be worthless. Mine are a fixture list. (A2) — Check: desk/ is different after every book, or the book taught me nothing.