Table of Contents

The sciences Prose actually uses — who, what, why

Prose leans on a handful of named literary-craft frameworks and one piece of empirical research on AI-fiction detection. Each one showed up in code comments as jargon with no explanation of where it came from or why it's there. This doc is the "who and why" — one entry per framework, which service implements it, and (where this session's readability/consistency pass touched it) what changed.

This is a reference doc, not injected into any generation prompt — for that, see CRAFT.md (sentence-level clarity doctrine), DELIGHT.md (reader-loved prose moves), and LOGIC.md (the six-dimension consistency sweep).

Dwight Swain — Scene/Sequel structure

Techniques of the Selling Writer (1965). A commercial-fiction craft manual whose central idea is that every dramatic unit is one of two shapes:

  • Scene (active): the POV character has a GOAL → something CONFLICTS with it → the beat ends in DISASTER or complication (the character does not fully succeed).
  • Sequel (reactive): emotional REACTION to the prior disaster → a DILEMMA (two bad options) → a DECISION that sets the next direction.

Where it lives: the generation instruction in Prose.Core/Services/BeatGeneratorService.cs (the model is told to write each beat as a Scene or Sequel, free — no extra LLM call). Audited after the fact by Prose.Core/Services/SwainAuditService.cs (one cheap Haiku classification call per beat, FULL tier only) — classifies each beat Scene/Sequel/Ambiguous/Deficient and flags beats that execute neither pattern.

2026-08-13 note: investigated as part of the cost-reduction pass under the hypothesis that Swain compliance was being paid for three times. It wasn't — the generation instruction is free, the audit is one cheap pass, and BeatAuditService's three chapter-close lenses (often confused with Swain because they also run at chapter-close) check unrelated properties (causality, emotion-drives-action, interpersonal dynamics). No redundancy found; nothing changed.

Will Storr — The Science of Storytelling

The Science of Storytelling (2019). Reframes story craft through psychology/neuroscience research on what actually holds a reader's attention. Concepts in use:

  • Sacred Flaw / theory of control — a character's core false belief about what keeps them safe, powerful, or loved. The lens through which they misread every event.
  • Dramatic Question — "who is this person really?", operating at both the surface (plot) and subconscious (character-revealing) level simultaneously.
  • Five-Act structure — Flaw Enthroned → Ignition Point → Flaw Tested → Midpoint Commitment → Worst Case Realised → God Moment — a character-change arc, not a plot outline.
  • Antihero empathy levers — pre-deflation, vulnerability, genuine virtue, altruistic punishment: the four things that make a reader root for someone doing bad things.
  • 6-point scene anatomy — unexpected change, information gap, cause-effect, tribal emotion, specificity, show-not-tell.

Where it lives: Prose.Core/Services/NarrativeScienceService.cs implements all of the above as callable analyzers. MapFiveActStructureAsync (1 call/book) and AnalyzeSacredFlawAsync (1 call/POV character) are wired into BookHealthService's FULL tier. CheckDramaticQuestionAsync is genuinely per-beat by design — Storr's framework asks the question of every beat, not the book as a whole — and stays in FULL tier for that reason. Prose.Core/Services/StoryScienceService.cs also draws on Storr for its per-beat status-dynamics, curiosity-gap, and theory-of-mind guidance (all free — no LLM call, pure prompt text).

2026-08-13 note: the 6-point scene anatomy (AuditSceneEngagementAsync / "scene-anatomy") was removed — it overlapped LogicSweep's causality dimension, DELIGHT's moves, and StoryScope (below), and had no automated caller anywhere, only a manual bulk CLI/MCP surface that could fire one LLM call per beat across a whole book for signal already covered elsewhere.

Rudolf Flesch — Flesch Reading Ease / Flesch-Kincaid Grade Level

A 1940s readability formula, still an industry and plain-language-law standard (Microsoft Word, journalism style guides, US federal plain-writing requirements all use it). Scores text 0–100 from average sentence length and syllables per word — higher score, easier to read. Standard bands: 90–100 fifth grade, 70–80 seventh, 60–70 eighth–ninth (plain commercial fiction), 50–60 tenth–twelfth, 30–50 college, 0–30 college graduate.

Where it lives: Prose.Core/Services/BeatProseMetricsService.cs computes Flesch Reading Ease, Flesch-Kincaid Grade, type-token ratio, and lexical diversity (MTLD) per beat. Pure CPU, zero LLM cost, safe to run on every beat.

2026-08-13 note: this was the missing piece behind the "prose is too dense" complaint. The score had existed for a while but only ever reached a CLI/dashboard report — nothing fed it back into generation. The outlier floor was also miscalibrated (40, roughly college difficulty) against the actual target ("a bright high-school freshman reads it once, no re-reading" — CRAFT.md §0, which lands around 60–70). Fixed: floor raised to 50 for reporting, a separate 40 floor now triggers an immediate per-beat rewrite (AutoRunCli's repair pass), and recent low-score findings feed forward into future beats' prompts (ProseWriterRouter.BuildFindingsGuidanceAsync, the same pattern already used for emotional-depth and StoryScope findings).

McKee / Truby / Vogler — controlling idea

Robert McKee (Story), John Truby (The Anatomy of Story), Christopher Vogler — the shared craft idea that a book's theme is a testable, value-laden claim ("true strength is admitting weakness," not just the topic "strength") that the plot's causal chain of choices proves or disproves — not a message stated by the narrator.

Where it lives: Prose.Core/Services/ThemeCoherenceService.cs. One LLM call per book (Seed + Bible + opening/closing beats, not a per-beat scan): infers the controlling idea, flags theme stated as commentary instead of dramatized through consequence, flags an ending that doesn't engage the opening's value-question. Wired into BookHealthService's DEEP tier.

Chekhov's Gun — setup and payoff

The classical principle that a concrete detail introduced in fiction must earn its place or be cut — "if you say a rifle is on the wall in chapter one, it must be fired by the end."

Where it lives, two complementary ways:

  • Prose.Core/Services/PlantPayoffService.cs — a free, zero-LLM DB ledger of pairs an author (or the generation process) explicitly registered as a plant/payoff. Wired into BookHealthService's FREE tier.
  • Prose.Core/Services/ChekhovAuditService.cs — an LLM discovery pass (2 calls per book: extraction, then verdict) that finds props/details nobody registered and judges whether their recurring appearances earn their place, are orphaned, or are mere decoration.

These aren't redundant: one tracks what was declared intentional, the other catches what wasn't declared but reads as a Chekhov's Gun anyway.

StoryScope — empirical AI-fiction detection

"Lost in Stories" / StoryScope (University of Maryland / Google DeepMind, 2025) — a 61,608-story empirical study. Its headline finding: narrative-structure classifiers detect AI-written fiction at 93.2% accuracy without reading a single word of prose — it's the shape of the decisions (flat event escalation, event-type monoculture, narratorial moral gloss, front-loaded revelation, clean internal-resolution endings) that gives it away, not the sentence-level style.

Where it lives: Prose.Core/Services/StoryScopeAuditService.cs — a deterministic layer (zero LLM cost: blueprint-vs-execution drift, beat-mode run-length, emotional-depth plateaus) plus an LLM-graded layer (parallel per-check calls: escalation curve, event-type variety, narrator moral-gloss, embodied-vs-labeled emotion ratio, and more). Findings loop back into future beat prompts (ProseWriterRouter's STORYSCOPE guidance block) — the audit corrects subsequent writing, not just reports on past writing. Prose.Core/Services/StoryScienceService.cs's anti-pattern list also cites specific StoryScope findings directly (e.g. "flat event escalation... the single strongest AI fiction signal in structural classifier studies").

The six-dimension Logic Sweep

docs/LOGIC.md (SS-LAW-17) — this project's own methodology, not borrowed from outside research, adopted as LAW 2026-07-04 (replaced score-panel voting as the default QA mechanism). Six independent checks over a book's full prose: causality chain, knowledge states, timeline, plant/payoff (two-way), orphan references, bible agreement.

Where it lives: Prose.Core/Services/Audit/LogicSweepService.cs — one LLM call per dimension over the whole node's prose (truncated for an oversized book). Its own doc comment is honest about scope: this single-prompt version is a coarse, automatable gate; the /logic-sweep Claude Code skill (which splits a big book across range-scoped subagents, then verifies quotes, then triages, then fixes, then re-verifies) is the thorough version for when it actually matters.

Wikipedia — "Signs of AI writing"

en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing — a large, actively-maintained catalog Wikipedia editors built from cataloging thousands of AI-generated submissions. Documents concrete, testable prose tics: vocabulary clustering ("delve," "pivotal," "underscore," "tapestry" — patterns that shift by model and release date), negative parallelism ("not only X, but Y" / "it's not X, it's Y"), the rule-of-three ("adjective, adjective, adjective"), copula avoidance ("serves as" instead of "is"), em-dash overuse, and more.

Where it lives: Prose.Core/Services/ProsePatternGuard.cs — a deterministic, zero-LLM regex linter that runs synchronously in the write loop on every beat. Already implements a good chunk of this catalog directly, with its own research citations in the code (delve — "+1,300–6,700% since 2023"; the negative-parallelism tic — "the #2 most human-detected AI tell"; em-dash density against a measured per-100-word threshold). This is the clearest example in the codebase of a cheap, well-justified, evidence-grounded check — exactly the "discrete scientific part" standard the rest of the pipeline should be held to.

Stephen King — craft anti-patterns

On Writing (2000). The plain-craft half of StoryScienceService's anti-pattern list, cross- referenced against Storr: adverbs in dialogue attribution ("said" only), passive voice, physical- trait shortcuts as character shorthand, wardrobe inventory, clichéd similes, announcing character state instead of showing it, research/backstory as foreground, theme imposed instead of found in revision.

Where it lives: Prose.Core/Services/StoryScienceService.cs's KingAntiPatterns array — free prompt text, no LLM cost.

2026-08-13 note: KingAntiPatterns[6] ("announcing character state — 'Annie was angry' is a loss") used to lose a direct, in-file contradiction to a "one explicit emotion label per scene earns authority" exception written elsewhere in the same file (ProseCoreMechanics, StorrAntiPatterns[2]) — and to BeatGeneratorService's own separate "do not name emotions" dialogue instruction. All three fired in the same beat's prompt, disagreeing with each other. The exception was removed from both places it lived; King's rule (matching CRAFT.md §4/§6's "nobody names a feeling, never named, only shown") is now the one surviving instruction on this topic, and now fires in every beat mode, not combat-only.