Surface Fundamentals · A quality floor for model-realized fiction

TIER C · CONTRACT  GENERIC FAILURE MECHANISMS  PROFILE-SENSITIVE THRESHOLDS

Surface Fundamentals are NovelForge’s default prose-realization safeguards. They exist because language models exhibit recurring cross-project failure mechanisms: fragmented pseudo-pace, camera-like micro-actions, narrator explanation, design-language leakage, empty embodiment, procedural reporting, fake significance, and other forms of polished text that do not behave like lived fiction.

Boundary ✦ Surface Fundamentals define generic model-failure mechanisms. They are not one novel’s house style, and passing them does not mean a chapter is engaging, emotionally effective, or production-ready.

01 · Where Surface Fundamentals sit

The rule stack is directional:

Framework Surface Fundamentals
→ Genre / Platform Profile
→ Project Profile
→ User Taste Profile
→ Current Request

Lower layers may tune thresholds, permit deliberate exceptions, or raise stylistic intensity. They may not silently convert a known cross-project model failure into the default behavior.

Clean prose is a floor. Reader Engagement is a separate positive-quality model.

02 · Realization principle: backstage design must disappear into lived fiction

Internal production concepts such as state transition, permission change, pressure, payoff, relationship delta, information advantage, resource constraint, and character arc should normally appear in prose as:

• people pursuing tasks;
• objects being used, lost, transferred, or misread;
• money, time, access, and obligations becoming binding;
• mistakes and incomplete information changing choices;
• refusals, bargains, interruptions, and consequences;
• relationship permissions shifting through behavior;
• characters noticing and judging according to their own knowledge.

If the narrator merely translates a Scene Card or state model into abstract explanation, realization has failed even when every sentence is grammatical.

03 · Production interface

Surface work begins after scene/story/character simulation has created something worth realizing.

A typical interface is:

bounded context + scene/character simulation
→ event-first Raw Draft
→ Surface Realization
→ deterministic / semantic surface checks
→ local rewrite or upstream regeneration
→ Reader Engagement / continuity / independent gates

Raw Draft is internal. Regression bad examples belong to post-generation checking and must not prime the first draft.

Surface repair must respect ownership:

• isolated sentence-level defect → local rewrite;
• repeated same-mechanism cluster → paragraph/block rewrite;
• multi-mechanism scene cluster → return to Scene Simulation;
• procedural flatness → causal scene design / Reader Pressure;
• functional-character collapse → Character Simulation;
• unowned interpretation → POV / semantic-ownership repair.

Do not “fix” surface safety by deleting all energy, humor, surprise, mystery, or rhetorical pleasure.

04 · Rhythm and segmentation failures

These mechanisms catch prose that simulates pace or cinematic importance through typography and shot selection rather than story movement.

HF-01 · NON-FUNCTIONAL FRAGMENTATION

Fail when sentence or paragraph fragmentation is used to imitate speed, seriousness, or cinematic cutting without a meaningful change in information, pressure, decision, ownership, or state.

Fast pacing should primarily come from faster information arrival, narrowing options, immediate opposition, deadlines, consequence, and state change—not from turning ordinary actions into isolated fragments.

A fragment-heavy project may explicitly opt in, but individual fragments still need narrative function.

HF-02 · MICRO-SHOT STORYBOARDING

Fail when one continuous action is decomposed into camera-like microshots that do not independently matter.

Repair by restoring the natural narrative unit. Do not merely join sentences mechanically if the underlying scene is still being narrated as a shot list.

HF-20 · SUBJECT / SENTENCE TEMPLATE REPETITION

Fail when consecutive sentences repeatedly restart with the same character name/pronoun + small action, or when a fixed short-short-long rhythm continues regardless of what the scene is doing.

Rhythm should follow information, action, voice, and pressure—not a reusable model cadence.

05 · Narrator overreach and abstract explanation

These mechanisms catch text that explains, advertises, or interprets more than the current narrative owner legitimately can.

HF-03 · NARRATOR THESIS / AUTHOR SUMMARY

Fail when the narrator explains the meaning that concrete scene evidence already established.

If deleting the summary preserves reader understanding, delete it. If the meaning is genuinely absent, repair event/choice/evidence before adding an abstract conclusion.

HF-04 · DESIGN-LANGUAGE LEAK

Fail when planning/database language appears in prose as narrator explanation without belonging to a character’s actual vocabulary.

Terms such as relationship upgrade, pressure node, permission delta, payoff, information advantage, or arc progression should normally become observable consequences rather than labels.

HF-07 · ABSTRACT EMOTION LABEL

Fail when vague labels such as “complex,” “indescribable,” “strange,” “confused,” or “moved” substitute for a concrete object of judgment, fear, desire, calculation, memory, or social management.

Useful interiority is specific about what is being tested or chosen.

HF-18 · ABSTRACT AGENT

Default fail when abstractions such as time, memory, history, fate, city, silence, darkness, or destiny are personified mainly to manufacture literary texture.

This is profile-sensitive. The question is whether the figure improves perception, voice, or meaning—not whether personification is globally banned.

HF-25 · EXPLANATION AFTER EVIDENCE

Fail when action/dialogue demonstrates a point and the narrator immediately restates the same point abstractly.

Trust the strongest layer. Do not duplicate meaning.

HF-27 · SEMANTIC ROLE MISATTRIBUTION

Fail when psychology, comparison, summary, or evaluative language belongs to the model/narrator rather than the POV or character who could truthfully own that wording now.

For every interpretive sentence ask: whose mind, knowledge, social position, and voice can legitimately produce this?

06 · Decorative significance and synthetic “writerliness”

These mechanisms catch prose whose main function is to look literary, cinematic, or important.

HF-06 · ORNAMENTAL METAPHOR

Fail when comparison or personification is primarily decorative, especially generic metaphors around body, time, memory, city, fate, silence, or history.

Literary profiles may permit more rhetoric. The invariant is function: perception, voice, meaning, or emotional precision must improve.

HF-15 · SIGNIFICANCE INFLATION

Fail when an ordinary object or action receives isolated emphasis, reaction pauses, contrast markers, or polished follow-up solely to make it appear meaningful.

Emphasis earns its space when the beat changes action, inference, risk, relationship, identity, location, resource, or another concrete state.

HF-16 · STAGED ROUTINE REVEAL

Fail when routine identity/context information is arranged as a synthetic movie revelation: inventory, narrow focus, isolated name/date/object, decorative detail, reaction pause.

Major irreversible information may deserve ceremony. Routine facts should usually emerge through purposeful action or search.

HF-19 · MANNERISM CONNECTOR

Words equivalent to “unexpectedly,” “instead,” “just then,” “as if,” “apparently,” or “it turned out” are not banned. Fail when they repeatedly manufacture contrast or significance that causality did not earn.

HF-29 · AI POLISH WITHOUT STORY FUNCTION

Catch-all cluster fail for sentences whose primary job is to sound polished, cinematic, profound, or “writerly” while adding no useful perception, voice, causality, tension, relationship, information, or rhythm.

Use a more specific HF code whenever one applies. HF-29 is not a license for vague dislike.

07 · Embodiment and object-use failures

Concrete detail is not automatically good. Embodiment must belong to task, space, agenda, and consequence.

HF-08 · EMPTY MICRO-ACTION

Fail when nodding, looking, touching a cup, rubbing fingers, pausing, or similar gestures are inserted only to make dialogue “visual.”

Keep an action when it changes timing, ownership, information, relationship, task progression, spatial constraint, or interpretation.

HF-09 · RANDOM EMBODIMENT PATCH

Fail when a disembodied scene is repaired by sprinkling unrelated gestures instead of restoring task, agenda, objects, space, and causal action.

HF-17 · PROP CATALOGUE

Fail when concrete detail is produced as inventory rather than selected by character purpose.

A prop earns space when someone uses, needs, pays for, moves, loses, transfers, inspects, hides, misreads, withholds, or makes a decision because of it.

08 · Character and dialogue failures

These mechanisms protect ownership, autonomy, and the continuing world around conversation.

HF-05 · DOSSIER INTRODUCTION

Fail when a character’s first appearance dumps age, clothing, job, history, personality, reputation, and attitude together without current scene need.

Reveal identity information when action, relationship, or interpretation uses it.

HF-10 · MECHANICAL DIALOGUE TAGGING

Fail when ambiguity is solved by attaching X said to nearly every turn while voice, agenda, task, knowledge, and spatial ownership remain absent.

Tags are tools. Ownership is the goal.

HF-11 · SPEAKER DRIFT / DISEMBODIED DIALOGUE

Fail when multi-character dialogue depends mainly on ABAB turn counting or line order.

Ownership may come from explicit address, distinctive agenda/voice, role-specific task/object, unique knowledge, spatial position, causal action, or another character’s response target. After a third speaker enters, reset alternation assumptions.

HF-12 · DIALOGUE WORLD ERASURE

Fail when extended conversation makes current work, objects, space, deadlines, and participant agendas disappear.

Dialogue happens inside a continuing world.

HF-13 · INTERVIEW / TRANSCRIPT DIALOGUE

Fail when conversation becomes pure information exchange with characters waiting for assigned lines.

Participants should pursue objectives, withhold, misunderstand, interrupt, bargain, test, evade, teach, refuse, or act while speaking when appropriate.

HF-30 · AGENDA-TO-DIALOGUE LEAKAGE / CHARACTER-SHEET-TO-DIALOGUE SERIALIZATION

Agenda drives speech; agenda is not speech. Fail when dialogue is a near-isomorphic natural-language serialization of a character’s private goal, role, risk, responsibility, knowledge boundary, fear, belief, unacceptable cost, or strategy rationale instead of a move shaped for the listener and the immediate situation.

The desired realization boundary is:

private character state
→ immediate tactic
→ listener model + shared context
→ social / relational cost
→ omission / compression / distortion
→ task or object interaction
→ utterance / action
→ interruption / response

Characters may be articulate, professional, rational, or long-winded. The semantic question is: Why is this person this complete, to this listener, at this moment, at this cost? A listener who already shares the meeting, history, role context, or obvious stakes usually does not need each speaker to restate a complete character sheet.

Do not diagnose HF-30 from line length, conjunction count, sentence fragments, ellipses, interruption count, colloquialness, or another lexical proxy. Complete speech is legitimate when completeness is itself the speech act or interaction requirement—for example testimony/deposition, incident debrief, board or military briefing, medical risk explanation, instruction/teaching, formal refusal rationale, clause negotiation, record-making, or an explicit confession/apology whose purpose requires a full account.

HF-30 is primarily an interaction/realization interface failure. Repeated or scene-level failure should route upstream to Character/Scene Simulation or writer-safe realization projection rather than be “fixed” by mechanically shortening every line.

HF-26 · FUNCTIONAL-CHARACTER COLLAPSE

Fail when a supporting character exists only to explain, praise, obstruct, deliver information, or trigger a scheduled beat.

Important characters retain agenda, work, information limits, emotional residue, relationships, self-interest, and plausible initiative.

09 · Constraint-defense failures

These mechanisms catch prose that visibly reveals backend prohibitions or tries to defend itself against imagined criticism.

HF-14 · CONSTRAINT LEAK / RULE DEFENSE

Fail when project/backend rules appear as negative proof inside prose—for example, the narrator explicitly explains why nobody suspected a forbidden trope simply because project rules prohibit it.

Ask: would this sentence exist if the backend rule had never been written? If not, realize the scene naturally.

HF-28 · CONTEXT DEFENSE PROSE

Fail when text explains why a character did not ask, notice, suspect, or behave according to an imagined trope solely to pre-empt criticism, unless that absence is itself causally relevant.

10 · Procedure and causality failures

These mechanisms are especially important in business, legal, political, technical, investigative, and competence-driven fiction where models tend to narrate correct process at equal weight.

HF-21 · PROCESS BROADCAST

Fail when routine operations receive equal narrative weight merely because the outline lists them.

Compress routine steps. Expand friction, error, disagreement, responsibility transfer, binding resource/time constraints, consequential choice, relationship movement, surprise, and consequence.

HF-22 · CHECKLIST CAUSALITY

Fail when a scene advances as a sequence of correct steps rather than consequences that create the next action.

Weak structure:

open → inspect → correct → check → send

Stronger causal structure:

problem → partial solution reveals cost → choice → reaction → changed options → consequence

Chronology is not causality.

11 · Suspense and ending failures

HF-23 · FAKE CLIFFHANGER / NARRATOR ADVERTISEMENT

Fail when chapter-end propulsion comes from abstract advertising such as “the real crisis had only begun” instead of changed story state.

Concrete consequence, incoming information/person/object, choice, reversal, relationship shift, voice, or next-state intrusion can provide forward pull without a mandatory twist.

HF-24 · FORCED MYSTERY

Fail when ordinary information is withheld or phrased vaguely only to create cheap suspense.

Mystery should arise from real information boundaries, uncertainty, deception, missing evidence, or character limitation.

12 · Paragraph, syntax, and POV fundamentals

Paragraphs are narrative units

A paragraph may combine action, observation, dialogue, response, local judgment, spatial/object change, and immediate consequence. It does not need all of them; it needs a coherent reason to end where it ends.

Standalone short paragraphs are strongest when they carry real interruption, irreversible action, key information, high-pressure pause, or directional dialogue.

Complete syntax by default

Commercial/readable prose normally prefers clear complete sentences. Fragments are available when they carry real semantic impact, not because “short = fast.”

Detail follows POV task

Concrete detail should be selected by what the focal character is doing, needing, fearing, comparing, searching, misunderstanding, deciding, or professionally trained to notice—not by an invisible camera inventory.

Narrator distance is purposeful

Narrator distance may vary by profile and POV, but the narrator should not behave like an editor outside the scene labeling significance and emotion after every beat.

13 · Diagnostic output

A surface audit should report observable mechanism evidence rather than private reasoning.

artifact_type: surface_audit
candidate_fingerprint:
findings:
  - mechanism_id: HF-XX
    severity: low|medium|high|cluster
    evidence_ref:
    scope: sentence|paragraph|block|scene
    repair_owner: surface|scene|character|reader|story
result: pass|rewrite|regenerate

A deterministic lint may flag candidate patterns. Final classification of nuanced prose mechanisms may require semantic judgment.

14 · Failure routing

Use the smallest repair layer that actually owns the defect:

isolated lexical/sentence hit        → local rewrite
repeated same-mechanism hits         → paragraph/block rewrite
multi-mechanism scene cluster        → Scene Simulation / regenerate scene
procedure/checklist flatness         → causal scene design + Reader Pressure
functional-character collapse       → Character Simulation
semantic-role drift                  → POV / character ownership repair
story premise/state failure          → Story / Canon layer

Do not patch a structural failure sentence by sentence until the prose is technically safe but lifeless.

15 · Invariants

1. Surface rules target mechanisms, not banned-word lists.
2. Profile-sensitive exceptions must remain intentional and functional.
3. Backstage design language should normally disappear into lived consequence.
4. Concrete detail must follow POV/task/causality rather than inventory.
5. Surface safety cannot substitute for Reader Engagement.
6. Regression bad examples remain post-generation evidence.
7. Cluster failure routes upstream rather than accumulating cosmetic patches.
8. A surface audit diagnoses prose; it never grants Canon authority.

16 · Related contracts

• Reader Engagement — positive quality after the prose clears the floor.
• Character & Relationship System — semantic ownership, speaker ownership, autonomy, and knowledge boundaries.
• Story System — causal scene design and planning responsibility.
• Quality Evolution — typed findings, revision routing, and candidate comparison.
• Production Pipeline — where Surface Realization and post-generation checking occur.