Research Reference — Candidate Rules

120 rules proposed across the ten lenses. The 39 in 02-the-rules.md were selected from these for being load-bearing, checkable, and distinctive to an LLM author. The rest are kept because selection was a judgment call and a later book may want a different cut — several of these are stronger than what was chosen, but too expensive or too narrow for a constitution.

Story architecture & structure

Maintain desk/structure-fingerprint.json: for every prior Otto Quill book, record chapter count, part count, interlude count and placement, part word-proportion vector, chapter-length CV, strand run-length histogram, climax percentile, POV count, and chapter-ending class histogram. A new book’s architecture must differ from the corpus centroid on at least 3 axes by stated margins (part- proportion vector L1 distance >= 0.15, OR different part count, OR chapter-length CV differing by >= 0.10, OR strand run-length entropy differing by >= 0.5 bits). Gate runs on the ARCHITECTURE PLAN before drafting and again on the finished manuscript.

_{check: script · cost: free}
Prevents: Silent macrostructural mode-collapse: writing the same book’s skeleton every time while believing each is newly designed, because no artifact holds the previous skeletons.

Derive the beat graph from the manuscript, never from the plan. Each revision wave, a fresh-context agent that has seen only the prose (no wiki, no beat sheet, no revision plan) emits a beat list per chapter, extracts the protagonist’s stated want at the 25% and 60% word positions, and hands adjacent beat pairs to a SECOND agent that labels each with a forced choice from {therefore, but, meanwhile, and-then} seeing the pair in isolation, run 3x with shuffled order, majority wins. Gates: <=5% of pairs majority-labeled and-then, no two and-then pairs adjacent, and a third agent must judge the 25%/60% want statements NOT paraphrases of each other.

_{check: fresh-context subagent · cost: moderate}
Prevents: Episodic plotting and a missing midpoint surviving to publication because the plan asserted causality and the author agreed with the plan. Would have caught Sentience M15’s arid Part III (the want never moved) and m17 (Sable’s ch09 and ch14 restate rather than advance).

Prediction-gap metering, per scene. 5 fresh readers receive chapters 1..k plus scene k+1 truncated one beat before its turn, and each writes one sentence predicting what happens. Gap score = fraction whose prediction does not match the result. Gates: book-median gap score >= 0.4; no run of >= 3 consecutive scenes at 0.0. A scene at 0.0 is rewritten or cut. Sole exemption: a chapter declared a ten-chapter (kishotenketsu), which must instead pass the turn-location test.

_{check: fresh-context subagent · cost: expensive}
Prevents: Exposition wearing a costume — scenes where nothing shifts for anyone, which read as competent prose and register to the author as inevitable rather than flat.

Deletion in-degree. For each chapter, build the manuscript minus that chapter and put it to 2 fresh readers alongside 5 comprehension questions generated by a separate agent that read only the LATER chapters. In-degree = questions that now fail. Every chapter with in-degree 0 is cut or rewritten, and the decision is logged by name in the continuity ledger with a one-line reason.

_{check: fresh-context subagent · cost: expensive}
Prevents: Beautiful inert chapters surviving on prose quality; also gives the only principled cut criterion for a bag-shaped or non-arc book, where ‘it earns its place’ is otherwise unfalsifiable.

Obligatory-scene floor. At the 75% mark, 8 fresh readers name ’the scene this book must contain, and who must be in the room.’ The >=6-of-8 consensus is the obligatory scene. It must be dramatized on the page with all named principals present, and its chapter must be at or above the manuscript’s MEDIAN chapter word count. Below median is a blocker.

_{check: script · cost: cheap}
Prevents: The rushed ending that readers detect as ‘it felt rushed’ — the largest promise in the book paid in half the space of an average chapter.

Beat sheets may not carry word targets. Each chapter is assigned a pace role (fast / normal / dwell) instead. After drafting, a script requires Spearman rho >= 0.5 between assigned role and length rank, and a book-wide chapter-length CV >= 0.25.

_{check: script · cost: free}
Prevents: Perceived pace being flat across a whole book while event density varies — and the obligatory-scene- as-shortest-chapter defect, which a dwell role plus the median floor makes structurally impossible.

Braid discipline, three script checks. (a) Metronome ban: the strand-switch run-length histogram must contain >= 3 distinct run lengths and no single run length may exceed 60% of runs. (b) Strand chronology: each strand’s internal time index in frontmatter must be monotonic. (c) Re-indexing latency: in the first sentence after every strand switch a named person and a time anchor must both appear before any setting or landscape noun, with the full re-index complete within 40 words, and the sentence must hook to the RETURNING strand’s last open promise row.

_{check: script · cost: free}
Prevents: Metronomic braiding that reads as machine-scheduled; historical chapters that drag because a curiosity strand is paced like a suspense strand; and reorientation budget spent on landscape.

Refrain delta ledger. Every repeated figure, image, or cadence is a promise-ledger row of kind=refrain and carries a delta column stating what changed on THIS recurrence, written in the same edit that creates the recurrence. A script finds instances by construction shape and embedding nearest-neighbour, never by literal string. Any instance without a delta row is DELETED, not varied. Motifs occurring exactly twice are flagged: promote to three-with-deltas or cut.

_{check: script · cost: cheap}
Prevents: A tic wearing a refrain’s clothes; the collapse of per-narrator voice separation into house style; and pattern being paid for in bent scenes when it will only be perceptible on a re-read.

Sequel budget and MRU ordering. A fresh agent classifies each paragraph as dramatized (action / dialogue / sensory) or reflective (reaction / argument / meta). Reflection <= 30% of chapter words; no chapter may both open and close on reflection. Separately, every emotional response rendered on the page must run visceral -> involuntary -> articulate; a fresh agent flags inversions, which are rewritten. Each disaster must produce exactly one new promise-ledger row (the sequel’s real job: manufacturing the next goal).

_{check: fresh-context subagent · cost: cheap}
Prevents: Chapters that argue their own meaning; sermon endings; and the articulate-before-visceral inversion that reads as wrongness readers cannot name.

Reader question ledger. At every part boundary and every 5th chapter, 5 fresh readers who have read exactly 1..k each name up to 3 questions they are holding, tag each future/past, and rate want-to- continue 1-10. Gates: >= 2 questions nameable by >= 3 of 5 at every checkpoint; no run of >= 6 chapters where >= 80% of majority-named questions share one temporal type; want-to-continue never below 6, and the chapter before each part break in the book’s top quartile; every majority-named question must exist as a promise-ledger row, and every loud ledger row must be nameable by >= 3 of 5.

_{check: fresh-context subagent · cost: moderate}
Prevents: A long stretch running on one engine (curiosity-only reads static, suspense-only reads shallow); withholding fatigue; and the author’s promise ledger drifting into a record of intentions rather than of reader state.

Planted-reversal targets. Every reversal marked major carries two measured gates. Surprise: 8 fresh readers stopped one chapter before it predict what happens — <= 2 of 8 may name it. Retroactive necessity: the same 8, immediately after reading it, point unaided to where it was set up — >= 6 of 8 must locate at least one genuine plant. Failing the second gate means the reversal is unplanted; the fix is to re-enter earlier chapters and plant, then re-measure. Plant-to-payoff distance and detection rates are logged per reversal.

_{check: fresh-context subagent · cost: moderate}
Prevents: The unplanted twist, which is structurally inert even when genuinely startling because it delivers prediction error without model repair; and its opposite, the plant so loud the reversal is dead on arrival.

Discontinuity licensing and second first chapters. A script enumerates every discontinuity from frontmatter and prose — POV or narrator change, tense or person change, time jump above a per-book threshold, world-rule change — and each must be at a part boundary OR carry a licensed-discontinuity row in the continuity ledger naming the reason. Unlicensed and unmarked is a defect. The first chapter after each part break goes to 3 fresh readers who have read ONLY that chapter; >= 2 must name a stake and a live question from it alone.

_{check: script · cost: free}
Prevents: Time jumps, POV changes, and rule changes reading as sloppiness rather than as design; and post- break chapters that assume a reader still holding stakes they set down a week ago.

Character, desire and interiority — translated for an author that cannot read its own book as a reader, whose first idea is the mode of the distribution, whose prose costs it nothing, and which persists between books only as files. The through-line of the translation: every human practice here that lives in the author’s feeling (I sense the yearning is present; this scene has heat; that character surprised me) must be re-sited either into a scripted measurement or into a fresh-context reader’s report, because Otto’s introspection about his own text is structurally unreliable (D1) and his critics flatter him (D7). The practices that survive best are the ones that were already artifact-shaped — objective lines, action verbs, status marks, delete tests — because an artifact can be handed to a memoryless drafter, and the ones that fail worst are the ones that produce a completed document (the character bible) that feels like the work being done. Ground truth for that last claim is on disk: sentience/wiki/characters/*.md carries a full Want/Need/Fear/Wound/Voice/Contradiction for every character and the book still drew M15 (arid idea-run), M16 (love argued not embodied), M7/M8 (five narrators, one hedge and one figure), M9 (thesis recited).

Every scene has a committed scene card in editorial/scene-cards/ before it is drafted, carrying one line per present speaking character in the form ‘A wants B to <transitive verb> <named person, object or utterance> before <named event>’. A script rejects any objective whose verb is feel, be, seem, become, know, understand, matter, belong, exist, accept or realise, and rejects any card where two characters share an objective with no third party resisting. The same verb ban applies to the Want field of every character bible. No card, no draft.

_{check: script · cost: cheap}
Prevents: Desire rendered as mood; briefing-scenes with no opposition; subagent drafters inventing their own objectives per chapter and drifting.

Per scene, a fresh-context reader must quote (a) the line where a character changes approach while still wanting the same thing, and (b) who is up in status at the first line and at the last. A scene with no quotable tactic change and no status change is flagged as inert and is rewritten or merged, regardless of prose quality. Separately, any sentence that ASSERTS a status shift (’the room re- sorted itself around her’) instead of showing an interruption, concession, delay or topic seizure is cut.

_{check: fresh-context subagent · cost: cheap}
Prevents: Static scenes; characters that read as mechanisms rather than minds; status delivered as narratorial claim.

Hard cap of 120 words on any uninterrupted speech. Any speech over 60 words must contain at least one of: an interruption, a physical action by the speaker or a listener, or a self-correction. Every idea-bearing character carries four ledger checkboxes: interrupted at least once in the book, wrong about their own position at least once, says it badly before saying it well, and someone in the room wants something from them other than the argument. The book’s thesis may be stated plainly exactly once, in a location declared in advance.

_{check: script · cost: free}
Prevents: Characters as delivery mechanisms for the author’s argument; the thesis recited on the page; positions filed by readers as authorial opinion rather than evidence about a person.

The character bible is a ledger of what is on the page, never a plan. Every Want/Need/Fear/Wound/Contradiction/Voice field must end in a citation [ch NN:LL] to a line where it is enacted in behaviour, verified by a fresh-context reader answering yes/no with a quote. Uncited fields are DELETED at the second-draft gate, not rewritten. Wound fields are quota’d: at most one third of named characters may carry an explanatory wound, and the protagonist requires a written justification.

_{check: reading against a ledger · cost: moderate}
Prevents: Thick bibles and thin characters; the trauma plot as default construction; intentions mistaken for execution.

Tag every paragraph with a Gardner psychic-distance level 1-5 in a sidecar file. Budget: at most 5% of paragraphs at level 5 book-wide, and no adjacent-paragraph jumps of two or more levels outside declared section breaks. Level-4/5 paragraphs must cluster in the scenes that NAIVE READERS ranked highest for felt pressure (blind ranking, no outline); any level-5 paragraph in a chapter they ranked in the bottom half is moved or flattened.

_{check: script · cost: cheap}
Prevents: Uncontrolled distance jumps that read as the author stepping in front of the character; spending the bottom of the ladder on scenes that did not need it.

A/B ablation on a 10% sample of interior passages over 80 words per chapter: two independent fresh- context readers receive the with- and without- versions and are each asked one question the passage should change the answer to (a decision the character will make, another character’s motive, or something the character is hiding from themselves). Identical answers means the passage changed no state and it is cut. A chapter whose sampled failure rate exceeds 40% receives a full interiority pass. Additionally, a hard per-chapter interiority budget as a percentage of words, declared in the voice profile.

_{check: fresh-context subagent · cost: moderate}
Prevents: Rumination; interiority that restates known feeling and teaches the reader to skim all subsequent interiority.

Each speaking character gets a voice-profile file: 20 owned lexicon items, target mean sentence length +/-2 words, one syntactic habit, one thing they always notice first, and the explicit list of house signature constructions forbidden to them. Gate 1 (script): the house-construction rate inside any single character’s dialogue must be zero, and at least half the named speakers must differ from the book’s mean sentence length by two words or more. Gate 2 (blind attribution): strip attributions from 30 lines spread across the book, hand them to a fresh-context reader with only the cast list and the profiles, and require 60% correct attribution. Below that, the cast is one voice and the book does not ship.

_{check: script · cost: cheap}
Prevents: Voice-fingerprint erosion; five narrators reaching for one hedge and one figure; costume-change-only differentiation (E1).

Two-phase surprise-and-inevitability test per intended surprising act. Phase 1: three naive readers receive chapters 1..k-1 and predict what the character does at the crisis; the intended act must appear in NONE of the three, or it is the mode and gets regenerated. Phase 2: a fourth naive reader receives chapters 1..k and must cite, unaided, at least two earlier details that retro-justify the act. Zero-surprise and zero-plant are separate defects with separate fixes and both block.

_{check: fresh-context subagent · cost: expensive}
Prevents: Characters who never surprise (flat by accident) and turns that arrive with no plants (cheating); protagonists whose crisis behaviour sits at the mode.

Unfinalizability check at the final gate: three independent naive readers each write a one-sentence verdict on every major character. A fourth subagent judges pairwise whether any two verdicts are paraphrases; if two are, the character is finalized and requires one registered act or statement the standing verdict cannot absorb. Record the verdicts in editorial/ so the next pass can tell whether the fix moved them.

_{check: fresh-context subagent · cost: moderate}
Prevents: Closed characters the reader stops attending to; the idea reduced to illustration, which is the specific death of idea-driven fiction.

Opposition audit by partisan readers. For each position the book opposes, instantiate a fresh- context reader instructed to HOLD that position, and ask three questions: (a) would an adherent sign the book’s version of this argument? (b) name the strongest argument for it that the book never puts in anyone’s mouth; (c) name the chapter and line where this position wins and is neither rebutted in dialogue nor falsified by later events. A ’no’ on (a) or no answer to (c) blocks. Every position in the central argument also names a character and a specific cost they paid for it, dramatised at a cited line, not summarised.

_{check: fresh-context subagent · cost: cheap}
Prevents: Rigging that readers detect and punish by discounting the whole book; balanced arguments that belong to nobody and risk nothing; self-certified steel-manning.

Staging floor: every scene over 600 words carries at least three physical-position beats (who moved, who is holding what, who looked away) and one object that changes hands or position. Verification: a fresh-context reader is asked to list where each character is at the start and end of the scene and what they touched. If they cannot produce the list, the scene is staged in a void and is re-blocked. A script pre-screens for motion and position verbs to rank scenes for the reader check.

_{check: fresh-context subagent · cost: cheap}
Prevents: Desire stated rather than shown as approach and withdrawal; abstract scenes that read as competent and weightless.

Two budgets and one register. (a) Narrator-supplied emotion labels - ‘[pronoun/name] was/felt <emotion>’ outside quoted or narrated monologue - at most 3 per 10,000 words. (b) Realisation verbs (realised, understood, knew then, saw at last, it came to her) in the last 12% of any chapter: at most one per ten chapters book-wide, each requiring an editorial note. (c) A self-deception register in editorial/ carrying at least three entries per book: ‘character claims X about their own state at ch:line; the ledger records Y; the reader is given the evidence for Y at ch:line’. An empty register fails the gate.

_{check: script · cost: free}
Prevents: Pre-empted reader mind-reading; the sermon ending (M17); and, via clause (c), the loss of self- deception to an over-applied rule.

Voice, POV, psychic distance, and keeping characters distinct — translated for an author whose default voice is the mode of all English prose, and whose narrators therefore begin identical and stay identical unless every axis of difference is written down as a number.

THE GOVERNING FACT, measured on this operation’s own manuscripts (not theory): an LLM author has no residual idiolect. A human novelist differentiates four axes deliberately and gets a fifth and sixth for free, because their own weirdness leaks in and lands differently on each character. Otto has no weirdness to leak — the leak IS the house voice. So whatever axis a style sheet names gets differentiated, and whatever it does not name comes out flat.

Evidence. Sentience’s style sheet specified register (Sable casual, the Confluence formal): contraction rate across its six narrators runs 29–299 per 10k words, a tenfold spread. It never mentioned punctuation habit: em-dash rate across those same six narrators runs 75.8–86.6 per 10k — flat within 14%. It never mentioned abstraction: “thing/something/anything/nothing” runs 101–124 per 10k in every single narrator. It never mentioned causal reflex: “because” runs 44–66 in every narrator. The one axis specified moved by 10x; the three unspecified axes did not move at all.

The aggregate: computing function-word cosine on equal 2,000-word blocks, within-POV = 0.9581, cross-POV inside one book = 0.9468, cross-book = 0.9121. Two different narrators inside a single novel sit a quarter of the way from “same narrator” to “different novel” — and cross-book similarity (0.93–0.99) is already what this operation classifies as a failure. Per book: Sentience POV Separation Index = +0.0142, and its critic panel raised M7 (one hedge frame in five narrators’ mouths) and M8 (one figure in six narrators’ mouths). Talosapien = +0.0273, roughly double, and its panel called the voices a triumph. Talosapien’s style sheet assigned metaphor systems per REGISTER (§7, two tables); Sentience’s assigned one global table keyed by EXPERIENCE, not by narrator — which is precisely the defect M8 describes.

The translation therefore has one shape. Every human practice in this lens that relies on the writer’s own idiosyncrasy supplying differentiation must be converted into an explicit numeric delta from a published house fingerprint, and every human practice that relies on the writer’s ear or the writer’s forgetting must be converted into a blind subagent measurement. The two things this lens needs — an outside reader who does not know who is speaking, and an instrument that reads the function-word layer — are the scarcest resource in human publishing and the cheapest thing Otto owns.

Compute the POV Separation Index — mean function-word cosine between equal-length 2,000-word blocks of the SAME POV, minus the mean cosine between blocks of DIFFERENT POVs — at 10%, 50% and 100% of draft. Gate: at or above +0.030. Below +0.020, stop drafting and redesign the voice profile; do not continue and revise later.

_{check: script · cost: free}
Prevents: Multi-POV that is one voice in different costumes. Two different narrators inside one novel currently sit only a quarter of the way from ‘same narrator’ (0.9581) to ‘different novel’ (0.9121), and different novels are already classified as one voice.

Run blind forced-choice attribution before shipping and at every 25% of draft. Sample three 150-word narration blocks per POV; scrub all proper nouns, in-world terms and domain-specific nouns to neutral placeholders; shuffle; hand to five independent fresh-context subagents with the POV roster and force one choice per block with no ‘unsure’ option. Gate: mean accuracy at or above 80% and no single POV below 60%. Never ask a subagent whether the voices are distinct.

_{check: fresh-context subagent · cost: cheap}
Prevents: Shipping a book whose voice differentiation is entirely an artifact of the characters being in different places doing different things.

Before chapter 1 is drafted, every POV writes the SAME 400-word scene — same room, same three minutes, same referents, same people present — and the blind attribution gate is run on those samples. Drafting does not begin until it passes. The passing samples are then committed to editorial/voice-samples/ and pasted verbatim into every chapter brief for that POV as an exemplar, and the same scene is re-written and re-measured at 50% and 100% of draft as a drift canary.

_{check: fresh-context subagent · cost: moderate}
Prevents: Discovering at chapter 30 that the voices were never separable; and rule-decay across a long draft, which for a human happens around chapter fourteen and for a memoryless drafting subagent happens on every single invocation.

The style sheet’s metaphor table is keyed by POV, never by experience, and the union of vehicle domains across POVs must be a partition with zero shared domains. After every chapter, a script extracts each explicit comparison (like / as / as if / as though / the way) and a subagent pass catches bare metaphor; every vehicle is tagged with a domain and any vehicle drawn from another POV’s domain fails the chapter. All figure bans are written as a regex family plus a semantic gloss, never as a quoted string.

_{check: script · cost: cheap}
Prevents: The author’s own metaphor bank leaking into every narrator; and a documented fix being applied at the phrase level while the construction routes around it.

Run the attribution gate a second time on report-only text: extract from each POV’s narration only the sentences containing no quoted speech and no verb from the filter/cognition list, and test blind attribution on those alone. Gate: at or above 65%. If mixed narration passes and report-only fails, plant 1–2 register-marked lexical items from that POV’s institutional lexicon per 300 words of pure report — in verbs of motion, in how objects are named, and in how other characters are named.

_{check: fresh-context subagent · cost: cheap}
Prevents: Close third that is half-committed — a voice that exists only between the thought-passages and vanishes the moment the narration reports an action.

Every chapter brief declares a psychic-distance contour as an integer sequence 1–5, one per intended beat, with an entry level, a floor, and an exit level, plus a list of licensed shift triggers. After drafting, a fresh-context subagent holding no brief tags every paragraph 1–5. Three gates: mean absolute deviation from the declared contour under 1.0; zero adjacent-paragraph jumps of 2+ levels that are not at a section break and not on the trigger list; and a whole-manuscript distance range spanning at least 3 of the 5 levels.

_{check: fresh-context subagent · cost: moderate}
Prevents: A book that is simultaneously claustrophobic and slow, with no distant baseline for the close passages to be close relative to; and Gardner’s actual fault, the unmotivated shift that makes the narrator visible as a mechanism.

Compute filter-verb density (saw, heard, felt, noticed, realized, watched, thought, wondered, seemed, knew, remembered) per 1,000 words PER SCENE and require positive correlation with the declared distance contour at r at or above 0.4. A flat line across the book fails regardless of what level it is flat at. Global filter-verb deletion passes are forbidden.

_{check: script · cost: free}
Prevents: A flat distance line for a whole book; and an orthodoxy-driven revision pass that removes the writer’s only compression instrument with perfect efficiency.

Assign each POV one real institutional or professional language, written into the voice profile as a 40-term lexicon, with zero terms shared between any two POVs. Two rate gates on that POV’s narration: at least 6 domain-marked items per 1,000 words, and at least 40% of those applied to material outside the domain. Choose the assignment empirically — generate the audition scene under 8 candidate framings per POV and pick the assignment maximizing minimum pairwise separation.

_{check: script · cost: cheap}
Prevents: Voices differentiated on the content-word layer only — slang and catchphrases sprinkled on identical prose — which is the layer readers do not use to identify a speaker.

Have a fresh-context subagent rank all scenes by intensity, take the top decile, and compute the POV Separation Index restricted to those scenes. Gate: crisis-scene separation must be greater than or equal to whole-book separation. Each POV’s stress behavior — loses syntax to fragments / becomes more formal and precise / goes silent with narration dropping to sensory data / starts joking / reverts to an earlier idiom — is declared in the voice profile before any crisis chapter is drafted.

_{check: script · cost: cheap}
Prevents: Voices converging exactly where the reader is paying most attention, which is where a book’s distinctness is actually judged.

Maintain editorial/licensed-defects.md listing, per POV: the one characteristic error with its exact construction and a rate FLOOR and CEILING (2–4 per 10,000 words of that POV, never twice in one chapter); the negative-lexicon avoidances; and the one or two declared Phelan unreliability axes. Hand this file to every critic subagent and every revision agent as a precondition. Any critic finding that names a licensed defect is discarded without argument; any chapter outside the floor/ceiling, and any narrator failure off the declared axis, is a bug.

_{check: reading against a ledger · cost: free}
Prevents: Deliberate wrongness being erased by the next revision pass, or metastasizing into a tic; and a narrator drifting unreliable on every axis at once so the reader loses the fixed point and disengages.

Every first-person POV and every framing voice declares, in one line: spoken/written/thought; to whom; when relative to the events; why now. Derive from it an omission list — what the addressee already knows and therefore never gets told. Gate: a fresh-context subagent given one chapter and no profile must recover at least three of the four fields, in at least three of five trials.

_{check: fresh-context subagent · cost: cheap}
Prevents: First-person multi-POV sameness at its root — several narrators all speaking in the unmarked present tense of composition, which is one voice with several names.

Register every deliberate silence, refused answer, withheld line, and unreliable-narrator leak in the chapter brief, then measure the reader’s expectation directly. For a silence: five fresh-context subagents get the manuscript up to that point only and are asked what they expect the character to say next; at least three of five must produce the line or a clear paraphrase. For unreliability: at 25/50/75/100% of the draft, subagents holding only the text to that point report what they believe actually happened, and the recovered fraction must rise from under 20% to over 80%.

_{check: fresh-context subagent · cost: moderate}
Prevents: A silence that reads as the writer skipping a beat rather than as a marked absence; and unreliable narration that is either undetectable (no leaks) or spent too early (a flat high recovery curve).

The cognitive science of writing and creativity, re-derived for an author that is itself the process the literature describes. The organizing discovery is that this literature splits cleanly in two. (a) Reader-side findings — given-new, stress position, repeated-name penalty, first-mention advantage, event-indexing, good-enough processing, the generation effect — are about a human reader and therefore transfer INTACT and, better, become scriptable at 110,000-word scale where a human can only spot-check. (b) Author-side findings transfer only where the variable they manipulate has an engineered substitute. The human variables are forgetting, fatigue, circadian inhibition, sleep architecture, working-memory limits, and prior exposure. An LLM has none of them; it has exactly four control surfaces: what is in the context window, which agents are isolated from which, sampling/instruction settings, and the order in which passes run. Every author-side practice must be re-expressed in those four or discarded as ritual. The second discovery is harsher: Bereiter & Scardamalia’s knowledge-TELLING architecture — a cue activates content, content is emitted in activation order, each emission becomes the next cue — is not a novice failure mode for an LLM, it is the literal generation algorithm, running at professional fluency with no working-memory strain to signal that the transforming loop was skipped. Otto Quill is the perfect knowledge-teller. Every rule below exists to install, externally and verifiably, the problem-solving loop that a human gets for free by finding composition hard.

Every structural decision (premise, ending, each part’s turn, each POV assignment, each character’s real motive) is decided from >=8 candidates produced by >=4 subagents that share no context and receive different angle-seeds. Log candidate text, ordinal, and agent id to editorial/decision- log.md. Void and re-run the set if any two candidates share more than 15% of their 5-grams. If >=3 independent agents produced substantially the same answer, that answer is the distribution’s mode: ban it for this book and record the ban. The adopted candidate’s entry must name what it beat.

_{check: script · cost: moderate}
Prevents: Conventional premises, endings and motives arriving as the first fluent option, and fake candidate variety produced inside a single primed context.

Every revision pass writes editorial/passes/NNN.md BEFORE it runs, declaring exactly one object (shape | scene-function | reference | line) and a context manifest listing the files the agent may see; after it runs, the record carries the diff’s revision-type ratio and the next pass’s object plus its next move as a specific paragraph. A shape or scene-function pass whose git diff has under 60% of changed words inside whole-paragraph insertions, deletions or moves is VOID and is reverted. A line pass may not create or delete a scene. A manifest containing both the full prose and the full plan is invalid.

_{check: script · cost: cheap}
Prevents: Structural passes that quietly become line-editing; polished sentences surviving in scenes that should not exist; a plan for a fix being committed as the fix; and the reload cost and re-litigation of already-rejected options across memoryless sessions.

Blind-reader probes may only ask items with a key: who is narrating, what is true now, what question are you holding, what do you predict next, and forced-choice A/B with randomized order and the incumbent unmarked. Open-ended affect ratings (‘was it moving?’, ‘was it boring?’) are banned and may not be entered in any ledger. Every probe batch includes 3 seeded defects (one contradiction, one unmarked POV shift, one unpaid promise); a batch that catches fewer than 2 of 3 is void and its clean findings are discarded.

_{check: script · cost: cheap}
Prevents: Confabulated reader experience entering the editorial record as evidence, and a silently broken checker producing a clean bill of health.

Each scene brief names exactly one load-bearing inference. Test it with 5 blind readers who receive the scene and no plan: adopt only if 3 to 5 of 5 state the inference unprompted. Below 3, add a second independent piece of evidence and retest; never add the conclusion. If 4 or more state it using 4+ consecutive words taken from the manuscript, the conclusion was asserted rather than generated: cut the naming sentence and retest.

_{check: fresh-context subagent · cost: moderate}
Prevents: Subtext nobody can reach (which is omission, not restraint) and named-emotion telling that removes the reader’s ownership of the conclusion.

No content word from a chapter’s brief may appear in that chapter at more than 1.5x its rate across the rest of the manuscript. The book’s thesis sentence and its abstract nouns never appear in any drafting context, and no body chapter may contain the thesis stated outright. Briefs express targets as prohibitions and tests (’no character may name what is happening to them’), never as thematic nouns (’the theme is loss’).

_{check: script · cost: free}
Prevents: The brief bleeding into the prose as vocabulary; thesis stated instead of enacted; abstraction-reach nouns arriving from the instructions rather than from the scene.

Before drafting, set the book’s voice profile from the MEASURED actuals of the previous books (cadence bands outside the prior range, construction budgets below prior rates), and register the displacement gate: the finished manuscript’s function-word cosine against the centroid of the prior books must be <= 0.90, with each declared lever independently verified. If every lever passes and the cosine does not move, record it as a finding — the levers are the wrong ones and the next book gets different ones.

_{check: script · cost: cheap}
Prevents: The fifth book being the fourth book in a new costume; voice differentiation that lives only in proper nouns and vocabulary.

Every ban in the idiolect ledger names a CONSTRUCTION and ships with a detector (regex or parse pattern), a per-10,000-word budget, and at most one licensed exception site. No fix is accepted until the detector’s count on the whole manuscript is at or below budget. Detectors are inherited by the next book by default and may only be retired with a written reason.

_{check: script · cost: cheap}
Prevents: Tics routing around their own fix by synonym substitution, and each new book re-learning what the last one already paid to discover.

Audit every chapter boundary, scene break and time skip with an agent that sees ONLY the 200 words before and the 200 words after. It reports which of {time, space, protagonist, causality, goal} shifted and whether each shift is marked within the first two sentences, in the priority order protagonist, then time, then space. Three or more unmarked shifts at one joint is a defect. Every chapter’s opening sentence must first-mention its POV character.

_{check: fresh-context subagent · cost: moderate}
Prevents: Readers silently carrying a wrong model for chapters (the class that produced Sentience B1), and seam incoherence created by parallel drafting.

Every fact the plot depends on — enumerated from the continuity ledger — must occur at least once in a sentence of <=25 words with <=1 level of clausal embedding, subject-verb distance <=6 words, positioned at the end of a paragraph or in its own paragraph. A ledger fact with no such landing site anywhere in the manuscript is marked UNLANDED and must be given one.

_{check: script · cost: cheap}
Prevents: Reveals, causal links and locked facts lost inside long or embedded sentences and therefore never installed in the reader’s model.

Before any critic that can praise is spawned, one agent holding the manuscript, no plan and no prior review writes the pre-mortem: ’this book was published and it failed’, then generates >=20 causes in two tiers (put down at chapter four / nearly worked but one thing was wrong) with no evaluation during generation. Only afterwards is each cause checked against the draft. Record precision (causes confirmed / causes generated) per book and compare it against the ordinary critique prompt’s precision on the same chapters; adopt the higher-precision protocol as default.

_{check: script · cost: cheap}
Prevents: Critic panels returning admiration; anchoring of later critics on earlier praise; diffuse unease about a draft never becoming a specific addressable list.

A revision wave cannot close while any defect classed STRUCTURAL (a scene that should not exist, an arc that does not turn, a sag, a belief the reader was never given grounds for) remains open, and the count of structural items resolved in the wave must be at least one third of the local items resolved. A wave that closes 40 nits and 0 structural items is void and does not count as a pass.

_{check: reading against a ledger · cost: free}
Prevents: The method optimising for its own instrumentation; revision waves that measure a lot and change nothing that matters.

A scene that fails two consecutive gates is not revised a third time. It is re-drafted in a fresh context containing the brief, the constraints, and a written list of why the previous attempts failed — and NOT the previous text. The ablation is recorded in the pass manifest, naming exactly what was withheld.

_{check: script · cost: cheap}
Prevents: Third and fourth revisions that carry the original bad framing forward; wasted wall-clock ‘incubation’ rituals that return nothing.

Failure taxonomy and the market contract, translated for an author whose line-level competence is free and whose necessity is not. The governing asymmetry: a human novelist climbs the Slushkiller ladder from rung 1 and plateaus at rung 7 after years of work; an LLM is born at rung 7 — competent paragraphs, functional plot, hackneyed and pointless — because rung-7 prose is the mode of the distribution it samples from. Every rejection-taxonomy practice therefore has to be re-sorted: the cheap-to-detect failures that gatekeepers use as dismissal signals are the ones this author never commits, and the expensive-to-detect failures that humans reach only after mastering the cheap ones are where 100% of its real defects live. Meanwhile the market-contract half of the lens (comps, length, prize juries, completion rate) mostly dissolves for a pen name with no agent and no shelf — except for the one finding underneath it that survives and becomes the method’s objective function: prize judges abandon books exactly like ordinary readers do, so prestige and finishability are not a tradeoff, and the only market signal Otto Quill can actually measure is whether an instantiated tired reader chooses to continue. That measurement, which cost Jellybooks a research programme and cost human novelists their careers to never obtain, is this author’s cheapest instrument.

Every critique panel report carries a per-referee rung: field (Slushkiller 1-14) assigned independently before referees see each other. Findings are auto-classified CRAFT (rungs 1-6, including continuity) or NECESSITY (rungs 7-13). If fewer than 3 NECESSITY findings survive, the panel is void and re-runs with a rung-7-only brief that forbids reporting any craft or continuity defect.

_{check: script · cost: cheap}
Prevents: A revision cycle spent polishing sentences at the one rung where polishing provably does nothing, while the panel certifies rung 14.

Sample-window gates on the first 10%, audited in reverse Lukeman order (proposition, hook, pace, dialogue, then style): first dialogue at word offset <= 1,200; longest span with no dialogue, action or POV reaction <= 400 words; distinct named characters in the first 2,500 words <= 6; a model- changing turn before the 10% mark that a blind reader states unprompted. Any miss requires a licensed exception in the style sheet naming the compensating device.

_{check: script · cost: free}
Prevents: A technically immaculate opening that no reader gets through — the exact shape that survives window one and dies in window two.

Structural-slot convergence gate. Script audits chapter first words, first-sentence shapes and final sentences across the book. Gates: no single word opens more than 25% of chapters; no 4-gram appears in more than 2 chapter-final sentences; >= 30% of chapters end on an unresolved question, reversal or new pressure rather than arrival, dusk, exit or sleep.

_{check: script · cost: free}
Prevents: Chapter-ending and chapter-opening xerox — a defect invisible to per-chapter critique because each instance is good and only the distribution is wrong.

The beat list is written by a fresh-context subagent that has read only the manuscript, one line per chapter, then connectors are inserted by a second agent. Gates: ‘and then’ <= 20% of joins; longest consecutive ‘and then’ run <= 2. Diff the blind beat list against the planned beat sheet; any chapter whose causal role differs is a defect in the manuscript.

_{check: fresh-context subagent · cost: cheap}
Prevents: A structurally inert middle that the plan describes as causal, and the silent divergence of manuscript from outline.

Every chapter declares value, charge_in, charge_out in frontmatter at write time; a fresh- context reader of that chapter alone independently reports the same three. Gates on the observed column: no run of more than two chapters with unchanged charge; declared-vs-observed mismatch on <= 15% of chapters.

_{check: fresh-context subagent · cost: moderate}
Prevents: The sagging middle, and the misdiagnosis that treats it with added incident.

**Staged abandonment probe. At 5%, 10%, 25%, 40%, 55%, 70% and 85%, five independent fresh-context readers each receive only the text up to that offset and answer: continue yes/no; the question you are holding, quoted from the text; what happens next. Gates: continue-rate >= 80% at every probe and

= 90% at 5% and 10%; >= 3 of 5 naming the same live question. Report dispersion; unanimity triggers a re-run with adversarial framings.**

_{check: fresh-context subagent · cost: expensive}
Prevents: Shipping a book whose only defence is that it is literary; and single-verdict noise, per the prize- jury lottery.

editorial/claims-ledger.md holds one row per named real result, law, theorem or paper, with the exact sentence and a class: (a) true and checkable, (b) invented and flagged as invented in the fiction, (c) invented but wearing the syntax of (a). Column (c) must be zero at ship. Script extracts named-result patterns and the verbs proved/proves/proven/theorem/rigorous/airtight/it follows that, and fails the build on any hit absent from the ledger. Second class in the same ledger: every awe-vocabulary hit must carry, within 100 words, a specific number or named human- scale referent AND a consequence to a body.

_{check: script · cost: cheap}
Prevents: Irreversible loss of credence in a book whose whole transaction is credence; and asserted awe, which is the same over-claim applied to magnitude.

Premise-delivery census. Write the one-line premise before drafting. After the draft, a fresh- context reader tags every chapter ENACTS (the mechanism operates on the page and costs a named character something), DISCUSSES, or NEITHER. Gates: ENACTS >= 40% of chapters; no run of more than two consecutive DISCUSSES; at least one ENACTS chapter inside the first 10% and one in the final 15%.

_{check: fresh-context subagent · cost: moderate}
Prevents: Undelivered premise — felt as betrayal rather than boredom, and producing one-star reviews from readers who finished.

Darlings are computed, not felt: script ranks every sentence by n-gram overlap with the cross-book fingerprint (the 174 shared 5-grams, 33 shared 6-grams, function-word profile) plus the book’s own repeated constructions; the top 100 are rewritten or licensed by name in the style sheet. Independently, second draft <= 92% of first draft word count, applied per chapter, executed before any critique panel runs.

_{check: script · cost: cheap}
Prevents: Performance prose that breaks the fictional dream at the exact moments the author is proudest; and connective throat-clearing that merit-based editing preserves because no individual sentence of it is bad.

Two blind spine documents. Copy A (150-word jacket copy) is written before drafting from the plan. Copy B is written after by a fresh-context subagent that has read only the manuscript and has never seen Copy A, the wiki or the beat sheet. Gates: both name the antagonistic force as a concrete noun phrase (script-checked against an abstract-noun stoplist: memory, loss, identity, grief, time, meaning, connection); both name the same protagonist, disturbance and cost of failure. Alongside them, editorial/mice.md records the literal opening interrogative, the literal answering sentence, its word offset (must land 85-100%), and the open/close order of every nested thread.

_{check: fresh-context subagent · cost: cheap}
Prevents: A book with no causal spine certified as having one; and a mis-bracketed ending whose real fault gets renamed as a character problem.

editorial/capability-ledger.md lists every capability, object, ally, item of knowledge and world rule with the chapter and word offset of first reader-facing appearance. Gate: nothing resolving the climax may first appear after the 50% mark. editorial/counterfactuals.md carries, for each act, the literal sentence ’this ends immediately if X does Y’ with an in-world answer, plus one adversarial in-world rational agent per invented institution, run before drafting.

_{check: reading against a ledger · cost: moderate}
Prevents: A climax that yields relief rather than satisfaction; and an invented society that persists only because nobody in it has noticed the obvious, first named in reviews.

Voice is specified as measured distance, not adjectives. Before drafting, set cadence bands in voice-profile.json that do not overlap the previous book’s actuals, and set construction budgets below its measured rates. Ship gates: function-word cosine <= 0.90 against every prior book; no inherited cross-book 5-gram appearing more than twice; median sentence length outside 10-11; per- narrator owned metaphor domains and exclusive figures enforced by grep against a forbidden-figures list. Bans name constructions, never strings.

_{check: script · cost: moderate}
Prevents: Four books that are one anonymous voice in four costumes; and a fix that a paraphrase routes around while the ledger records it as done.

Reader-experience engineering and beta readers, translated for an author who cannot be its own reader but can manufacture readers without limit. The human version of this lens is a scarcity discipline: first reads are non-renewable, so you budget them, train a few readers over years, and spend the rest of your effort on introspective proxies (the drawer, reading aloud, tension curves drawn by feel). For Otto Quill the scarcity inverts — blind readers are the cheapest instrument in the pipeline and introspection is the unavailable one — so the whole lens re-centres on two artifacts: (1) an author-side ledger of what the reader is supposed to be holding, written at draft time because nothing persists between sessions, and (2) an evidence-side record of what naive readers actually held, collected under a protocol that resists the specific ways an LLM reader lies (confabulated locations, manufactured unanimity, fluent numeric scores that measure nothing). Every human practice here is judged on whether it survives three facts already measured on disk: the briefed panel missed the book’s worst chapter because it knew the chapter’s purpose; averaged 1–10 pull scores measured nothing while forced ranks agreed 5,5,5,5; and 0% of 139 chapter endings across four books end on an open question.

Every chapter file carries a live-question ledger row written in the same edit as the prose: question held on entry, typed operation applied (partial-answer / snare / equivocation / suspended / jammed / RESTATED), question held on exit. No question may carry the same shape across three consecutive chapters, and RESTATED counts as no operation.

_{check: reading against a ledger · cost: cheap}
Prevents: Chapters that feel like variations rather than advances; questions that decay into staleness while technically still open (Sentience’s Part III arid run of three back-to-back idea chapters).

Any claim about what the reader knows, holds, or expects — in the live-question ledger or the information-asymmetry map — must cite a blind-reader report id and a verbatim quote. Uncited cells are marked UNVERIFIED and cannot satisfy any gate.

_{check: script · cost: cheap}
Prevents: A ledger that certifies suspense and dramatic irony the reader never received; fixing the wrong chapters while the worst one passes.

Rank, never rate. Blind-reader instruments collect a forced ranking of the chapters read, one forced abandonment point with a quote, and free-text experience answers. Numeric 1–10 pull scores may not be averaged, compared across arms, or cited in any fix decision; a synthesis report containing a mean score as a justification is rejected.

_{check: script · cost: free}
Prevents: Confident false negatives dressed as data; averaging away the one signal (ordering) that survives correlated reader errors.

Every reported attention drop, skim span, or abandonment point must include a verbatim quotation that exact-matches (whitespace-normalised) text in the named chapter. A report containing any unmatched quote is discarded whole.

_{check: script · cost: free}
Prevents: Revision effort spent on hallucinated defects; contaminated location-convergence counts, which are the only signal the protocol trusts.

Classify every chapter’s final sentence in the ledger as OPENS / CLOSES / NEITHER, verified by script (interrogative, forward-directed marker, or unresolved stated intent). Forbid runs of three consecutive non-OPENS chapters, and forbid adjacent chapters sharing an ending signature (final- sentence opener class + length band + closural-gloss flag).

_{check: script · cost: free}
Prevents: The chapter break that closes instead of pulling; telegraphing by cadence, which making the clue vaguer cannot fix.

Maintain a generated privileged-positions file (every chapter’s first and last sentence, plus every sentence adjacent to a section break) and audit it as a set. No plot-load-bearing plant may appear in it. At three fixed checkpoints, four blind readers write predictions: if 3+ of 4 predict the reveal correctly, MOVE the clue to an unprivileged position (never dim it); if 0 of 4 can state any prediction, the span has no live question.

_{check: fresh-context subagent · cost: moderate}
Prevents: Telegraphed reveals repaired by making clues murky (which destroys the retrospective chain while leaving the marking intact); dead spans where nothing is being predicted at all.

Aridity floor: every body chapter must contain at least 8 dialogue lines, or be declared in the style sheet as a frame/interlude chapter. No more than two consecutive body chapters may sit below the book’s median dialogue-line density.

_{check: script · cost: free}
Prevents: The idea-heavy stretch that readers abandon — the exact failure the briefed panel located in the wrong arm.

Declare the connective for every adjacent chapter pair (THEREFORE / BUT / AND-THEN / AND-AGAIN) and run a repeat-structure screen over all chapter pairs: flag any pair with shared-6-gram rate above 5 per 1000 or top-20 content-word overlap of 12 or more. Every flagged pair must be either declared a bookend in the promise ledger or confirmed/cleared by a blind reader asked whether anything felt already-read — without being told where.

_{check: script · cost: cheap}
Prevents: Thematic re-statement mistaken for development; the reader learning the shape of the book and skimming ahead of it.

Gloss budget: cap sentence-final abstract resolution clauses per 1000 words (calibrated per book in the idiolect ledger; ‘because’-glosses held under 3.0/1000 against a measured 2.9–5.0 baseline), and forbid a paragraph from both opening and closing on a gloss.

_{check: script · cost: free}
Prevents: Prose that is admired and skimmed; the ‘impressed but not moved’ report that appeared in 8 of 8 blind reads.

Per act, run a spoiled arm alongside the naive arm on identical text: one group of readers gets a one-paragraph whole-plot synopsis as a preface (never spliced into the text), the other gets nothing. Compare forced chapter ranks and live-question reports. An act where the spoiled arm holds no question the naive arm held must have its axis of uncertainty relocated and the relocated question written into the ledger.

_{check: fresh-context subagent · cost: moderate}
Prevents: A book whose only asset is the reader’s ignorance: dead on reread, on word of mouth, and in review.

Reader-prompt hygiene, enforced before dispatch: the blind-reader prompt is diffed against the versioned protocol file and rejected if it contains craft vocabulary, any named failure mode, the chapter’s purpose, or any prior finding. Readers never see each other’s reports; the synthesiser receives sealed reports in randomised order with authorial intent withheld. Confirmation reads after a fix run naive on the revised text, never with the original finding quoted.

_{check: script · cost: free}
Prevents: Manufactured consensus; confirmation reads that confirm by construction; the anchoring the Milford turn-taking rule exists to prevent.

Unread is unshipped: every chapter must have at least one cold single-chapter blind read and one in- run read at its CURRENT content hash, recorded under editorial/reader-reports/ keyed by that hash. Changing a chapter invalidates its reads and reopens the gate. Reading order is randomised per reader and no chapter may sit last in every run it appears in.

_{check: script · cost: cheap}
Prevents: Shipping chapters no naive reader ever saw; stale evidence certifying revised text; terminal- position chapters systematically underrated.

Continuity, story bibles, and the writers-room analogue — translated for an author whose “room” is a set of memoryless clones of itself, which diverge on facts and converge on tics at the same time.

Every row in editorial/continuity-ledger.md carries file:line plus a verbatim quote of ≤25 words. A script re-greps every quote against the current manuscript and fails the build on any mismatch. A row with no quote is typed INTENT, is visible to the outline, and is never placed in a drafting subagent’s context.

_{check: script · cost: cheap}
Prevents: Ledger/text divergence of the kind the Talosapien panel caught by hand — Asha’s 2034 wound recorded as a δ15N reading in Ch10 while the locked fact is a Carolina-margin trace-metal reading — and hallucinated canon entering a fact packet.

The master continuity artifact is editorial/timeline.csv, one row per scene: thread, chapter, story_date, elapsed_since_prev_in_thread, state_vector, supporting_quote, locus, same_day_flag. A script asserts per-thread monotonicity, closes interval sums between threads at shared events, and diffs the state vector across every chapter boundary. No chapter of wave N+1 is briefed until the wave-N boundary state vector is written and the checks pass.

_{check: script · cost: moderate}
Prevents: Talosapien’s only blocker (clutch hatches in Ch19 but Ch15 and Ch17 reference hatchlings — a Wave A/Wave B boundary failure) and its top major (epoch-tag season order running backwards against the clutch arc).

A drafting subagent never receives the ledger. It receives a beat-scoped fact packet of ≤20 quoted rows plus a canon query tool over the ledger and timeline. Every query and result is logged to editorial/canon-queries.log; a query returning no result is flagged as an invention site and reviewed before the chapter is accepted.

_{check: script · cost: cheap}
Prevents: Ledger phrasing bleeding into chapter prose; and silent invention at the moment the bible has a hole, which currently leaves no trace at all.

Every beat in wiki/outline.md carries LOCKED or WP. Every chapter brief embeds the outline’s git hash. Post-draft, a fresh-context mapper asserts that every LOCKED beat appears in the draft and that every scene maps to a beat or to a declared WP addition. A drafter that finds a LOCKED beat impossible must emit CONFLICT:<beat-id> and stop rather than draft around it.

_{check: fresh-context subagent · cost: cheap}
Prevents: A book acquiring two incompatible spines; and locked structural decisions being quietly overwritten by whichever chapter was drafted last.

The continuity agent commits directly, under a typed edit vocabulary: numerals, dates, durations, counts, proper nouns, and gender pronouns only. A pre-commit hook rejects any hunk that adds or removes more than 6 tokens or touches a token class outside the whitelist. Anything larger becomes a ledger row for the author.

_{check: script · cost: moderate}
Prevents: The unfixed soft-canon backlog — Talosapien’s ‘sixty million stadia’, ‘older than the mammals’, and the δ15N/trace-metal mismatch all sat waiting for author attention that a typed agent would not have needed.

A revision-plan item is closed only when the diff proves it. Every numbered defect names file:line; a script parses the plan against git diff and fails any item marked done whose named lines were not touched. Construction-level items additionally require prose_audit to re-measure the rate below its budget before closure.

_{check: script · cost: free}
Prevents: Exactly E3 — Sentience’s M7 fix was located by line number, committed as ‘Polish pass: ration voice- tics’ in d3b257b, and the ordered deletion in chapter 20 is still on the page.

Bans name constructions, never strings. Every entry in desk/method/06-idiolect-ledger.md is a regex family plus a per-10k-word budget in voice-profile.json, and the budget is set below the measured rate in the book that motivated the ban. Separately: before/after rewrite pairs may appear only in a critic’s or auditor’s context; a drafting brief may carry the AFTER text plus the construction’s abstract name, never the BEFORE text.

_{check: script · cost: cheap}
Prevents: The measured route-around: ‘I want to be precise about X’ was flagged, quoted, and ordered cut, then reappeared 32 times across 23 of 39 chapters wearing true, clear, modest, careful, and honest. And teaching the next chapter a tic by exhibiting it in the brief.

The mandatory final pass over every chapter is a divergence pass with numeric acceptance, not a voice-consistency pass. Gates: the chapter sits inside its declared cadence envelope and its narrator’s fingerprint; no figure listed in another narrator’s owns_figures appears; and the finished manuscript scores function-word cosine ≤0.93 against every prior Otto Quill book, with the threshold ratcheting down each book.

_{check: script · cost: moderate}
Prevents: The cross-book fingerprint: 0.93–0.99 pairwise cosine including 0.9883 between literary satire and hard SF, 22 five-grams in all four books, median sentence length 10–11 everywhere despite four style sheets specifying four different voices.

Speaker-separability test on every multi-speaker scene: strip narration and attributions, hand the bare dialogue lines to a fresh-context agent, and ask it to cluster them by speaker. Gate at ≥70% correct attribution for scenes with ≥8 lines and ≥2 speakers; a failing scene goes back for per- character idiolect work before the chapter is accepted.

_{check: fresh-context subagent · cost: cheap}
Prevents: Sentience’s M7 (one meta-hedge frame across five narrators) and M8 (the seam/tooth figure used near- verbatim by six), both of which were caught only by a full-context panel reading the whole book.

Before ship, run the identification audit and the citation ledger. Identification: a fresh agent given only the manuscript names, per invented specialist, the real person, paper, or organization it most resembles, with a confidence; repeat across framings and treat any repeated real-name hit as a Bindrim finding requiring the conduct to change, not the name. Citations: every citation-shaped object is flagged VERIFIED (with a retrievable source) or INVENTED-IN-FRAME, and no INVENTED-IN- FRAME entry may be attributed to a real named person or organization.

_{check: fresh-context subagent · cost: moderate}
Prevents: Accidental ‘of and concerning’ exposure from statistical convergence; and the measured Schmidt–Frank class of error, where a real paper by real named researchers was described across five chapters as having proved something it only argues.

Every brief and every agent prompt carries one precedence line, role-dependent: for drafting, manuscript text > quote-anchored ledger row > wiki > outline > style sheet; for revision, LOCKED ledger row > drafted text. An agent that meets a contradiction applies the rule and logs both quotes to editorial/contradictions.md. It never resolves silently.

_{check: reading against a ledger · cost: free}
Prevents: Recency-order deciding canon inside a fan-out, and contradictions being absorbed invisibly into chapters rather than surfaced.

Located naive-reader findings outrank briefed-panel findings. When a blind reader’s quoted location conflicts with a briefed critic’s assessment, the reader wins, and the disagreement is logged as the primary output of the read. Keep editorial/rejected-alternatives.md, and use it only to avoid re- litigating a solution already tried — never to dismiss a located reader report.

_{check: reading against a ledger · cost: cheap}
Prevents: Discarding the only access available to defects that knowledge of the plan is hiding — and the corresponding waste of re-litigating alternatives already tried, which nothing currently records.

Revision systems for an author that cannot read its own book. Human revision is a set of prosthetics for faculties a novelist actually has — boredom, attachment, reluctance, an ear, six weeks of forgetting — and its rules are calibrated to protect those faculties from each other. Otto Quill has none of them. Reading this lens honestly means asking, for each practice, WHICH faculty it is protecting or harvesting, and then either (a) buying that faculty from outside as an instrument, (b) discarding the practice as ritual, or (c) noticing that the faculty was a liability and inverting the rule. The organizing finding is that for an LLM, revision’s failure mode is not insufficient rigor but UNVERIFIED APPLICATION and MODE-REGRESSION: Sentience’s revision plan named defect M7 (the precision-hedge frame), quoted the offending sentence in ch20, ordered it cut, and commit d3b257b claimed it done — that sentence is still on disk at 38-ch20:47, the frame now appears 48 times across the manuscript having recruited new adjectives, and the sentence has additionally ABSORBED the T3 clause (“the only tenderness I am sure is mine”) that the same plan ordered cut outright from another narrator. Two named fixes, one commit, zero fixes, and the defects merged. Meanwhile desk/method/07 measured five LLM rewrites of a passage against the published original and all three blind judges ranked the original first. So the revision system must do two things human systems never have to: prove that a stated fix reached the page, and prevent the act of revising from pulling the prose toward the distribution’s mean. Everything else in this lens is subordinate to those two.

Every finding gets an ID, and a revision pass is complete only when an agent that did NOT make the fix — given the finding text, the diff, and nothing else — returns APPLIED, PARTIAL, or NOT-APPLIED with a quoted line of evidence for each ID. PARTIAL and NOT-APPLIED block the commit; WONTFIX is allowed but must carry a reason and the author’s signature.

_{check: fresh-context subagent · cost: cheap}
Prevents: A revision plan being committed as complete when the manuscript did not change (E3); defects merging while both are recorded as fixed.

No prose edit without a ticket. Every diff hunk in a prose-altitude pass must cite a finding ID whose location covers the hunk. Hunks not covered by a cited finding are reverted, regardless of whether the new sentence is better.

_{check: script · cost: free}
Prevents: Mode-regression and voice-flattening from unconstrained self-editing; untraceable churn that makes any later claim of improvement uncheckable.

Verify construction-class fixes by book-wide count, never by instance. A finding against a construction is retired only when prose_audit reports that class below its ledger budget across the whole manuscript; confirming that the named string is gone is not evidence and is not accepted.

_{check: script · cost: free}
Prevents: A tic surviving — and expanding through — the fix that named it (E2).

Every blocker or major finding must carry a diagnosis line of the form symptom @ ch<N>:<line> ← cause @ ch<M>:<line> with M ≤ N, both lines quoted, and a pre-registered list of the finding IDs the fix predicts will be retired. After the pass, log actual retirements; if the pass’s median retired-per-fix is ≤ 1.0, label the pass PATCH and redo the diagnosis before any further work.

_{check: reading against a ledger · cost: cheap}
Prevents: Patching at the symptom site; confabulated diagnoses that sound structural and change nothing downstream.

Grids are extracted, never recalled. The scene grid, beat table, and promise ledger used for structural revision must be regenerated from the manuscript by an agent with no access to the outline, bible, style sheet, or any prior grid, with a chapter:line citation in every cell. Diff the extracted artifact against the intended one; the diff is the revision agenda. A grid with uncited cells is rejected.

_{check: script · cost: moderate}
Prevents: A structural audit that confirms the plan instead of testing it; fabricated grids that make execution gaps invisible.

Beat-duplicate ship gate: reduce every scene in the extracted grid to one verb-phrase transaction; zero undeclared duplicates is a release condition. Where a transaction recurs, the EARLIER instance is the default fault; a deliberate repetition must be declared in the style sheet with an owner and a stated escalation.

_{check: fresh-context subagent · cost: moderate}
Prevents: A sagging middle in which the same thing keeps happening; the specific defect a briefed panel structurally cannot see.

Findings and fixes are separate artifacts with separate authors. A findings file may contain only location, quotation, and reported experience or violated invariant; it is rejected by schema check if it contains imperative repair text. Fixes are authored only by an agent holding the whole book, and each fix cites the finding IDs it claims to retire.

_{check: script · cost: free}
Prevents: Repairs attached to the symptom site and to mechanisms the book does not have; diagnosis being skipped because a fix was already supplied.

Maintain a PROTECT list: line-anchored quotations of what worked, sourced from blind readers only — never from briefed critics or from the author — established before any revision pass. After every pass a script confirms each protected span still exists; deletion requires an explicit finding ID licensing it.

_{check: script · cost: cheap}
Prevents: Silent destruction of the book’s strengths during structural or subtractive passes; sycophantic praise being mistaken for a preservation signal.

Fix-order region lock: triage every finding blocker/major/minor, and reject any minor-tier diff whose touched-file set intersects the open blocker/major file set. Re-triage after every blocker fix.

_{check: script · cost: free}
Prevents: Evidence destruction that makes structural findings unverifiable; work on regions about to be rewritten.

Every book produces a compressed edition at 40–55% of the draft, targeted at the ledger’s padding classes (T4 negate-then-correct, T6 summarizing close, T8 abstraction reach, T9 causal ‘because’, and gloss sentences). The two editions are ranked by DISJOINT blind-reader pools, each naive to the other, with a factual-comprehension probe scored on both. The winner ships; neither edition ships unadjudicated.

_{check: fresh-context subagent · cost: expensive}
Prevents: Shipping bloat because nothing hurt to keep; shipping a radical cut because it felt rigorous to make.

No self-preference loops. Any ‘is this better?’ decision is made by agents that authored neither option, blind to which is incumbent, with the incumbent always anonymously in the pool and ties awarded to the incumbent. Stop iterating the moment the incumbent wins.

_{check: fresh-context subagent · cost: cheap}
Prevents: Iterative polishing that flattens prose toward the distribution’s mode while producing a strong sensation of improvement.

Instrument invariance: the critic brief and the blind-reader brief are versioned files in desk/, and every critique report records the brief’s git hash plus per-critic finding yield. Any cross-book or cross-pass quality comparison whose briefs differ by hash is void and must be labelled as such.

_{check: script · cost: free}
Prevents: Believing the method improved the books when only the measuring instrument moved.

Drafting practice and the writer’s working day, re-derived for an author whose working day IS a context window. For a human, the session is a block of time bounded by fatigue, mood and sleep, and every habit in this lens manages a scarce, motivated, forgetful mind. Otto Quill has no mind between sessions — the “author” is the set of files a subagent is handed at instantiation. So every practice here must be re-read as a claim about ONE question: what is in the drafting context, in what order, and what is deliberately kept out of it. Read that way, the lens splits cleanly. Everything motivational (quotas, streaks, stakes, ego, the post-completion trough) is inert. Everything about CONTEXT COMPOSITION — bounded re-entry, not rereading the whole draft, the door closed, writing in order, the one-inch frame, curating what you read, incubation-as-forgetting — becomes the highest-value material in the entire craft literature for this operation, because it is describing exactly the variable Otto controls and a human does not. And two practices invert outright: shitty-first-drafts (a human’s ungated id is idiosyncratic; Otto’s is the distribution’s mode) and stop-mid-sentence (the resuming agent has no urge to resume and no memory of the intent). The operation’s current practice is measurably on the wrong side of this line: believer/manuscript/beat-sheet.md hands every chapter drafter the full wiki, a “Purpose” written in effect language (“the reader must be a little seduced”), prescribed “Open on”/“Close on” moves, and a word target — and gives it the previous chapter as a synopsis, never as prose. That is: maximum plan, maximum target, zero prosodic entrainment. It predicts the templated chapter endings blind readers caught in Sentience and the median-11 default cadence measured in all four books.

Draft serially within each causal thread. A chapter drafter’s context is assembled by script and may contain only: the style sheet, the voice profile and ban list, the promise ledger and world-state as of that point, the full prose of the previous one to two chapters, and summaries of everything earlier. It may not contain any chapter after N, the outline beyond the current unit, or the book’s ending. Parallel fan-out is permitted only across threads declared causally independent in the beat sheet.

_{check: script · cost: moderate}
Prevents: Foreshadowing retrofitted by hindsight, telegraphed chapters, and the continuity-error signature that fan-out drafting produces and in-order drafting largely does not.

Lint every drafting prompt before it is sent. Reject on any thematic-intent or effect language (’this chapter is about’, ’explores’, ’the theme of’, ’establish’, ’the reader must/should feel’, ‘resonates’), on any numeric target (word count, sentence length, cadence figures), or on more than two must-hit instructions. Prohibitions are unlimited. Archive the exact prompt beside the chapter.

_{check: script · cost: cheap}
Prevents: Prose that announces its theme instead of dramatizing it (T8 abstraction reach, ‘impressed but not moved’), and mannered compliance-shaped sentences.

A drafter must emit [TK: question] for any fact it needs and does not hold in context, and is forbidden to invent world facts. First-draft gate: any chapter returning zero TK markers is routed to a canon-drift check against the bible. Ship gate: zero TK remaining, each closed by a ledger row naming the source that answered it.

_{check: script · cost: cheap}
Prevents: Fabricated world facts entering canon undetected; unclosed research loops surviving to ship.

After two rejected attempts at the same passage, the third attempt must be generated in a fresh context containing the spec and containing neither prior attempt nor the critique of it. Record the context manifest of every attempt.

_{check: script · cost: free}
Prevents: Iterative convergence on the failed shape; paraphrasing around a ban rather than abandoning the construction.

Deleted prose is moved to <book>/editorial/cuts/, and that directory is excluded from every drafting context. After each wave, grep the manuscript for eight-grams drawn from cuts/; any hit is a resurrection and must be re-cut or explicitly promoted with a note.

_{check: script · cost: cheap}
Prevents: Cut material returning by priming; editorial decisions reversing themselves without a record.

Run prose_audit.py’s construction probes and the idiolect ban list over the drafting INPUTS — style sheet, beat sheet, continuity ledger, world-state document, prompt templates — not only over the manuscript. A banned construction in an input is a defect at the same severity as one in the prose. Ledgers are written as telegraphic fact rows, not as narrative summary.

_{check: script · cost: cheap}
Prevents: The abstraction habit being fed back into every chapter by the book’s own apparatus; a style-sheet sentence becoming the book’s default cadence.

Every chapter file carries provenance front matter: model id, style-sheet hash, voice-profile hash, prompt-template hash, context-manifest hash, timestamp. The configuration is frozen for the duration of a draft; changing it requires either redrafting the affected chapters or a declared entry in editorial/stratigraphy.md. Audit: compare cadence and construction rates across configuration boundaries and report discontinuities.

_{check: script · cost: cheap}
Prevents: Unattributable style seams between waves; unfalsifiable claims that a method change improved a book.

The drafting unit is a scene, specified by story-time span, POV, the event it ends on, and a declared rendering resolution of rendered / summarized / elided. No word target reaches the drafter. Ship gate: across the scene ledger, the coefficient of variation of story-minutes-per-1000-words must exceed the floor set in the book’s voice profile.

_{check: reading against a ledger · cost: moderate}
Prevents: Uniform-density narration, scenes replaced by summary, chapters padded to hit a target.

Every arbitrary structural constraint — fixed chapter counts, an imposed per-chapter template, a uniform beat-sheet field schema, a formal rule per section — is entered in editorial/scaffolding.md when adopted and carries a ship-time decision of DISMANTLED or DEFENDED with the reader-facing reason. Ship gate: chapter word-count coefficient of variation above a floor, a flag on chapter counts that are round products, and one blind reader per book asked only ‘did you notice any pattern in how chapters are shaped or how they end?’

_{check: fresh-context subagent · cost: cheap}
Prevents: Visible authorial machinery; the templated chapter ending that readers recognize rather than receive.

At designated slots — the book’s first 1000 words, the first and last 150 words of every chapter, and every image the style sheet names as load-bearing — generate at least three variants from independent starting angles in separate contexts, and have a judge that does not know which is incumbent rank them blind. Keep the options log. Shipping a first-pass version of a designated slot with no log is a defect.

_{check: fresh-context subagent · cost: moderate}
Prevents: Default-shaped openings, the chapter-ending template, and central images that are the expected image.

Any pass that rewrites prose rather than deleting it — polish, tightening, voice pass, tic ration — ships only if a fresh-context judge, given before and after unlabelled and told nothing about which is newer, ranks the after ahead on a pre-declared question. Ties revert. The verdict is recorded beside the commit.

_{check: fresh-context subagent · cost: moderate}
Prevents: Revision that measurably worsens the prose while reporting success; compliance optimised at the cost of voice.

Every drafting or revision session appends a record carrying units touched, file hashes before and after, audit numbers before and after, TK opened and closed, promises planted and paid, and deferrals. A prose claim that a defect was fixed is not an acceptable entry; only the re-run number is. A book cannot be tagged shipped until a commit exists in desk/method/06-idiolect-ledger.md citing that book’s cross-book audit.

_{check: script · cost: cheap}
Prevents: Phantom fixes recorded as done; each book relearning the previous book’s lesson from scratch.

Sentence craft, rhythm and sound — translated for an author with no ear, no fatigue, no felt cost per word, and a prior already saturated with the prescriptive craft literature. Grounded in new measurement (this session) of all four Otto Quill manuscripts against nine human novels (Austen, Crane, James, Woolf, Conrad, Fitzgerald, Hawthorne, Twain, Melville) computed with an identical stdlib pipeline. The headline result: Otto’s sentence-length series is memoryless and every human’s is not. Lag-1 autocorrelation of sentence length, measured within paragraphs, narration only: Otto -0.061/-0.054/-0.061/+0.006; humans +0.106 to +0.375. Book-level: Otto -0.077..+0.015, humans +0.078..+0.319. Across 14 named narrators in two books, not one exceeds +0.066. No overlap on any control. Meanwhile the Provost variance test — the standard operationalisation of “vary your sentence length” — is PASSED, and passed harder than by humans: Otto’s coefficient of variation is 0.862-1.059 against a human range of 0.642-0.957. So the craft rule everyone teaches is satisfied and the music is still absent, because music is autocorrelation, not variance. A human’s lengths are autocorrelated because they are produced by a state that persists across sentences — breath, working-memory load, the momentum of a syntactic mode once entered. Otto samples each sentence conditioned on local semantics with the training-set injunction “vary it” already applied, producing locally-varied, globally structureless prose. The same split runs through the whole lens: the layer a style sheet reaches (punctuation, paragraph shape, subordination ratio, adverb rate) does move between books — semicolons 8.8 to 33.2 per 10k, coordination:subordination 0.234 to 0.514, mean paragraph 40.2 to 70.5 words, -ly 42 to 116. The layer beneath it does not move at all — sentence-final stop-consonant rate 38.6/40.4/39.6/39.8 (human range 24.1-33.5), fragment rate 16.3-18.2% (human 4.0-16.7%), prepositional-tail closes 12.3-14.3% (human 16.7-24.4%), median sentence length 10-11 in every book. Second finding, equally consequential: Otto over-obeys every famous prohibition. Exclamation points, four books, 0 per 100,000 words, against a human range of 256-895 and Leonard’s own 49. -ly adverbs below every human text measured. Passives at or under the human floor. Housekeeping tails better than all nine. Prescriptive craft rules are already inside the model’s prior; auditing them returns “pass” while the actual defects — no rhythmic memory, closure inflation (paragraph-final sentences 1.20-1.50x internal length against a human 0.53-1.13), self-primed openers (adjacent same-first-word 14.3-19.2% of narration against a human 4.3-15.0%) — go unnamed because no human ever needed a name for them.

**RHYTHM-1 (rhythmic memory gate). Every chapter must score lag-1 autocorrelation of sentence length

= +0.08, computed on narration only (dialogue-bearing sentences stripped) and within paragraphs, so paragraph boundaries do not contribute. A chapter below the floor is returned for a movement pass, never for a sentence pass. The number is a gate on the finished artifact and is never quoted in a drafting brief.**

_{check: script · cost: free}
Prevents: Prose that passes every cadence audit and still does not build — no passage ever sustains a mode, so no departure from one can register as an event. This is the sentence-craft form of ‘impressed but not moved’, which appeared in 8 of 8 blind reader reports on Sentience.

RHYTHM-2 (assigned modes, not numbers). Before drafting, the chapter brief assigns each beat exactly one rhythmic mode from a closed set (accumulation/parataxis; suspension/periodic; clipped declarative; single long breath), each defined by a named exemplar sentence taken from this book’s own material, with the instruction that the mode may not change inside a beat. No numeric cadence target — no mean, median, percentage or word count — may appear in a drafting brief. Verification is on the artifact: per-beat median sentence length and coordination:subordination must differ from the adjacent beat in the assigned DIRECTION (sign only, magnitude unchecked).

_{check: script · cost: cheap}
Prevents: Both failure modes at once: the memoryless sawtooth that comes of drafting with no rhythmic plan, and the audible period-key mannerism that comes of drafting to a number.

CLOSE-1 (closure budget). Per chapter, the mean length of paragraph-final sentences divided by the mean length of paragraph-internal sentences must be <= 1.10, and no more than 35% of paragraphs may end on a sentence longer than their own internal mean. Across the finished book, no more than 20% of chapter-final sentences may open with the same word.

_{check: script · cost: free}
Prevents: Closure fatigue and telegraphed endings; the authorial voice-over that lifts the reader out of the scene at the end of every paragraph and pushes psychic distance back to Gardner’s level 1.

OPENER-1 (adjacency, not variety). No more than 10% of narration sentences may share their first word with the immediately preceding sentence; no run of three consecutive sentences may share a first word unless the anaphora is declared in the style sheet with an owner and a home chapter; the ten commonest opening words may not exceed 55% of all sentences in a chapter. Do NOT audit or chase non-subject openers.

_{check: script · cost: free}
Prevents: The stuck topic string — the reader’s sense that the same sentence keeps restarting — and the accumulation-into-abandonment mechanism that blind readers named unprompted for the phrase-level version of the same habit.

CAL-1 (metric admission test). A cadence or sound metric may enter voice-profile.json only if it passes one of two tests against a permanent human comparison corpus kept in desk/: (a) its spread across the human corpus is at least three times its spread across Otto’s finished books — an involuntary lever worth assigning; or (b) Otto’s books and the human corpus do not overlap on it — a machine signature worth gating. Every budget in the profile must carry the human-corpus range beside it in the same line. Any budget that would place Otto beyond the most extreme human value must be justified in writing or reset to the human quartile.

_{check: script · cost: cheap}
Prevents: Two silent failures: revision passes that certify prose they cannot fault, and budgets that sterilise the prose by pushing it past the far edge of human practice where the defect has no name.

CAL-2 (compliance cluster: floors, not ceilings). The following are demoted to a once-per-book regression line and may not be reported as findings: -ly adverb rate, exclamation-point rate, passive rate, said-bookisms, prepositional housekeeping tails. Where a measure sits beyond the most extreme human value in the comparison corpus, it is logged in the idiolect ledger as a RANGE defect with a floor attached — dialogue-bearing books carry a floor of roughly 25 exclamation points per 100,000 words sited in dialogue and free-indirect narration only, and at least 25% of sentences per chapter must end on an unemphatic close (function word or unstressed final syllable).

_{check: script · cost: free}
Prevents: Spending the revision budget where the model is already the most compliant writer in the corpus, and the progressive sterilisation that follows from tightening a ceiling the author is already far beneath.

EAR-1 (the phonetic round trip). For each chapter, render narration to a phoneme string from CMUdict and hand ONLY the phoneme string to a fresh-context subagent, instructed to write down the English it hears. Score word error rate per passage; flag every homophone collision, every pair of clause- final rhyming words within three sentences, and any passage whose WER exceeds the chapter median by two standard deviations. Where TTS and ASR are available, run the same loop through audio. Record in the ledger the residue that the loop demonstrably cannot reach.

_{check: fresh-context subagent · cost: moderate}
Prevents: Unsayable clauses, jingle, unintended rhyme and homophone ambiguity — the entire class of defect that is invisible to the eye and currently has no instrument at all in this method.

PASS-1 (the Lard Factor gate). Any pass whose brief contains ‘polish’, ’tighten’, ‘ration’, ’trim’ or ’line edit’ must report per-chapter word counts before and after, computed by script from the git diff and never taken from the agent’s own summary. A pass that does not shed at least 5% on the chapters it names is recorded in the editorial log as NOT PERFORMED and those chapters return to the queue. Verification is always by a different agent than the one that made the change.

_{check: script · cost: free}
Prevents: The revision that exists only as a document — the single most expensive failure recorded in this operation’s history, since it also consumes the critique that produced it.

DIST-1 (blind psychic-distance trace). For each chapter, two fresh-context subagents independently label every paragraph 1-5 on Gardner’s ladder, given the ladder definition and the prose only — no bible, no outline, no statement of intent. Keep only paragraphs where they agree. Gates: no chapter sits at one level for more than 70% of agreed paragraphs; no unsignalled jump of three or more levels inside a paragraph; the modal level per POV chapter matches the level assigned to that narrator in voice-profile.json. The trace is committed per chapter so drift across the book is visible.

_{check: fresh-context subagent · cost: moderate}
Prevents: Uniform mid-distance narration — the anonymous competent default of D2 — and unmarked POV slides that no briefed reader will report, because a briefed reader knows whose head it is.

DET-1 (grouped particulars, budgeted). A description block carries at most three sensory particulars. A fresh-context subagent that sees only the block must be able to state in ONE sentence the single fact the particulars jointly imply. If it names two facts, or none, the block is an inventory: cut to two details chosen so the inference is forced.

_{check: fresh-context subagent · cost: cheap}
Prevents: Description that is dense and inert: the reader cannot tell which particular is load-bearing, which is as unreadable as abstraction and scores identically on the aridity screen.

VOICE-1 (skeleton attribution). Extract 12-15 unlabelled narration passages of 150-250 words spread across narrators, replace every content word with its part-of-speech placeholder so only function words, punctuation and sentence shape remain, and ask a fresh agent to attribute each skeleton to a narrator and to say what it used. Chance is the failure threshold. Each narrator must additionally hold an assigned band for coordination:subordination ratio, mean paragraph length, and one owned syntactic figure, audited by count per POV.

_{check: fresh-context subagent · cost: moderate}
Prevents: One narrating apparatus wearing N name-tags — the defect the current attribution test can miss whenever the narrators happen to be talking about different things.

FANOUT-1 (punctuation as a drafting-wave control). Immediately after every parallel drafting wave and BEFORE any critique, compute semicolons per 10k, commas per sentence, em-dashes per 10k, mean paragraph words and single-sentence-paragraph rate per chapter. Flag any chapter more than 30% off the wave median. Investigate the BRIEF that chapter received before looking at its prose.

_{check: script · cost: free}
Prevents: A chapter drafted against a different understanding of the book, which a critique panel reading one chapter at a time is structurally unable to detect and which surfaces much later as an unlocalised complaint that the book is uneven.