Expand framework for literary fiction, deeper anti-patterns, and organic finishing

[SHL0MS]

PR: Expand framework for literary fiction, deeper anti-patterns, and organic finishing

Summary

Used the autonovel framework to write a 22,000-word literary novella (Provenance — contemporary fiction, DFW-inspired voice, about an aging conceptual artist who clones himself as his final work). Then read the framework's first shipped novel (The Second Son of the House of Bells). Both experiences exposed gaps: patterns that survive all existing checks, genre assumptions that don't generalize, duplicated content across files, and a missing phase.

This PR adds genre-agnostic literary fiction support, 7 new anti-patterns derived from actual output analysis, an organic finishing phase, a prose critique methodology, and deletes 123 lines of duplicated content.

+579 / -123 across 5 files.

What's in here

New anti-patterns (from reading Second Son)

The novel passes every word-level slop check and still reads as AI within three paragraphs. These are the structural patterns that survived:

• §14 The Type-Description Formula: "the expression of a man who had been disappointed by rooms full of people before" — turns every character into an instance of a sociological type instead of a person. The #͏1 AI character-description tell. Not in any existing list.
• §15 Scene-Structure Repetition: Every scene follows arrive→observe→interact→reflect→leave. Paragraph-template (§1) applied at scene scale.
• §16 Universal Competence: When every character in the cast is high-competence, the world becomes a seminar. Nobody fumbles, panics, or says something stupid.
• §17 Magic System Re-Explanation: The under-note sensation gets the full sensory description in every scene. Trust the reader after the first demonstration.
• §12 One Emotional Temperature: Every scene processed through the same analytical register. The narrator never loses composure.
• §13 Thesis-Absorption: Every detail connects to the acoustic premise. Nothing in the world resists the concept.

New anti-slop patterns (from writing Provenance)

• Paired em-dash parentheticals: The — X — mid-sentence construction, with frequency budgets (~3 dashes per 1000 words total)
• "Which was" recontextualization machine: Density thresholds (max ~2.5 per 1000 words)
• Deflationary sentence endings: "...which was refreshing" — a few is voice, a pattern is a tic

Literary fiction support (CRAFT.md §9)

• Milieu as constraint system (a real-world professional setting functions like a magic system)
• The Dangerous Insight Test (insider knowingness vs. genuinely uncomfortable truth)
• Plotless and recursive structures (question, accretion, recursive) as alternatives to beat sheets
• Literary fiction evaluation rubric

Phase 2.5: Organic Finishing (program.md)

A new phase between revision and done. The agent rereads and adds details that resist the story's logic: unexplained sensory details, sideways memories, physical comedy, unexplained character behavior, unresolvable recurring details. Distributed unevenly. Purpose: make the story exceed its argument.

Character and dialogue craft (CRAFT.md §2)

• Supporting Character Opacity Test: Four checks for whether a non-POV character is a person or a function (independent desire, counter-narrative action, wrongness/selfishness, unreadability)
• Dialogue Balance and Rhythm: Checks for line-length variation, asymmetry, dead air, and imperfection

Prose critique methodology (program.md)

Structured prose critique to supplement numerical scoring: emotional temperature, character opacity, passivity audit, surprise audit, register variation, dramatization audit. Output is a list of specific problems, not a number.

Stability trap — concrete protocols (CRAFT.md §7)

The existing section correctly diagnoses the problem but prescribes only abstract advice. Added three checkable protocols:

• Register break requirement (mandatory for works >10K words)
• The unforced decision (protagonist must make at least one genuine choice)
• The scene that breaks the rules (at least one moment the preceding material makes impossible)

Voice template expansion (voice.md Part 2)

• Rhetorical Tic Budget (voice-specific construction caps with examples for two voice types)
• Punctuation Discipline (em dash budget, paired-dash rules, semicolon/colon guidance)
• Register Variation (dominant mode + counter-mode + deployment triggers)

Deletions (-123 lines)

• voice.md Part 1: 90 lines of banned word tables and structural patterns that duplicated ANTI-SLOP.md verbatim. Replaced with 6-line cross-reference.
• program.md Stability Trap: 9 lines of abstract advice duplicating CRAFT.md §7. Replaced with cross-reference.
• CRAFT.md §5: 6-line "detectable telling patterns" list duplicating §6's fiction AI tells.
• ANTI-PATTERNS.md §6: "Section Break as Rhythm Crutch" (subsumed by new Scene-Structure Repetition at the correct scale). Renumbered §7-§18.

What's NOT changed

• The core pipeline architecture (foundation → draft → revision loop)
• The evaluation scoring system (augmented with prose critique, not replaced)
• The fantasy-specific content in CRAFT.md (preserved — literary additions are additive)
• The existing anti-slop word lists (all retained)
• The existing 11 anti-patterns (all retained, one merged)

Context

These changes come from two sources:

1. Writing with the framework: Provenance, a 22K-word novella using CRAFT, ANTI-SLOP, and ANTI-PATTERNS as guardrails, with iterative critique-revision cycles that exposed gaps in voice calibration, rhetorical tic detection, and organic detail
2. Reading the framework's output: The Second Son of the House of Bells, which passes all existing checks and still reads as AI due to structural patterns (type-description formula, scene repetition, universal competence, magic re-explanation) not covered by any existing anti-pattern

The thresholds (e.g., max 2 "which was" per 1000 words) are calibrated against these two data points. They should be validated against additional output and adjusted.