Voice Calibration System

Discovering Your Narrative Voice Through AI Competition

Overview

The Voice Calibration System is Phase 2B of the Writer’s Journey. It uses a multi-model tournament to help writers discover and lock in their unique authorial voice before entering the drafting phase.

Core Philosophy: Rather than the writer trying to describe their voice abstractly, they see 15-25 writing samples from different AI models and simply pick the one that sounds most like them.

Where It Fits

┌─────────────────────────────────────────────────────────────────┐
│                    THE WRITER'S JOURNEY                          │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  Phase 0    Phase 1      Phase 2B        Phase 3      Phase 4   │
│  Research → ARCHITECT → VOICE MODE → DIRECTOR MODE → EDITOR    │
│  (NotebookLM) (Story Bible) (Tournament)  (Drafting)   (Polish) │
│                              ▲                                   │
│                              │                                   │
│                         YOU ARE HERE                            │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Prerequisites: Story Bible complete (ARCHITECT Mode) Unlocks: DIRECTOR Mode (scene drafting with voice consistency)

The Tournament Process

1. Configuration (SETUP)

The writer provides:

Test Prompt: A scene description for agents to write
Test Context: Setting, mood, characters, stakes
Voice Description (optional): Hints about desired style
Agent Selection: Which AI models to compete (3+ required)

2. Generation (RUNNING)

Each selected agent generates 5 variants using different strategies:

Strategy	Focus	Characteristics
ACTION_EMPHASIS	Physical conflict	Fast pacing, external action, dialogue in motion
CHARACTER_DEPTH	Psychology	Slower pacing, internal landscape, emotional truth
DIALOGUE_FOCUS	Conversation	Subtext, conflict through words, relationship dynamics
BRAINSTORMING	Exploration	Multiple perspectives, experimental approaches
BALANCED	Hybrid	Mix of elements, standard structure

Example: 4 agents × 5 strategies = 20 variants generated concurrently (~2-5 minutes)

3. Selection (REVIEW)

The writer reviews variants in a matrix view:

Rows: AI Agents (Claude, GPT-4o, DeepSeek, etc.)
Columns: Writing Strategies
Features: Preview, word count, optional scores, multi-select for comparison

4. Configuration (SELECT WINNER)

After picking the winning variant:

POV: First person / Third limited / Third omniscient
Tense: Past / Present
Voice Type: Narrator / Character voice / Hybrid
Metaphor Domains: Boxing, weather, architecture, etc.
Anti-Patterns: Phrases to avoid
Phase Evolution: How voice changes through story acts

5. Bundle Generation (COMPLETE)

Four reference files are created:

File	Purpose
`Voice-Gold-Standard.md`	Reference sample + authentication tests
`Voice-Anti-Pattern-Sheet.md`	Patterns to avoid, AI tells
`Phase-Evolution-Guide.md`	How voice adapts by story phase
`voice_settings.yaml`	Scoring weights and thresholds

How It Integrates With GraphRAG

The Voice Calibration system produces artifacts that enhance GraphRAG context injection:

Voice Bundle → Scene Generation

When DIRECTOR Mode generates a scene:

┌─────────────────────────────────────────────────────────────────┐
│                    SCENE GENERATION CONTEXT                      │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  From GraphRAG:                   From Voice Bundle:            │
│  ├── Character ego-graphs         ├── Gold Standard sample      │
│  ├── Active HINDERS edges         ├── Anti-patterns list        │
│  ├── Unresolved FORESHADOWS       ├── Metaphor domains          │
│  ├── Current tension score        ├── Phase-appropriate voice   │
│  └── Pacing recommendations       └── Scoring thresholds        │
│                                                                 │
│                         ↓                                       │
│              LLM generates scene with:                          │
│              - Narrative consistency (Graph)                    │
│              - Voice consistency (Bundle)                       │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Verification Integration

The Tiered Verification System (GraphRAG Phase 4) can check voice:

Tier	Voice Checks
FAST	Anti-pattern detection (forbidden phrases)
MEDIUM	POV/tense consistency, metaphor usage
SLOW	Full voice authenticity scoring via LLM

How It Integrates With Narrative Dashboard

The Tension Indicator and Pacing Dashboard (from NARRATIVE_DASHBOARD_UI.md) provide feedback that influences voice:

Phase Evolution Trigger

// When tension changes significantly, voice should adapt
if (currentTension > 0.7 && previousTension < 0.5) {
  // High tension phase: use "climax" voice profile
  voiceProfile = phaseEvolution.climax;
}

Pacing-Voice Alignment

Pacing Type	Recommended Voice Characteristics
`fast`	Short sentences, action verbs, minimal introspection
`slow`	Longer sentences, more internal landscape, description
`balanced`	Mix of above
`concluding`	Resolution tone, emotional payoff, callbacks

The Narrative Dashboard can suggest voice adjustments based on pacing analysis.

Architecture

Backend

backend/
├── services/
│   └── voice_calibration_service.py    # Tournament orchestration
├── models/
│   └── tournament.py                   # Data models
├── agents/
│   └── orchestrator.py                 # SceneTournament, DraftCritic
└── api.py                              # Endpoints

Frontend

frontend/src/lib/
├── components/
│   ├── VoiceCalibration.svelte         # Master component (TO CREATE)
│   ├── VoiceTournamentLauncher.svelte  # Configuration UI
│   ├── VoiceVariantGrid.svelte         # Matrix display
│   ├── VoiceComparisonView.svelte      # Side-by-side compare
│   ├── VoiceBundleGenerator.svelte     # Bundle generation
│   ├── VoiceVariantSelector.svelte     # Winner selection
│   └── VoiceEvolutionChart.svelte      # Evolution tracking
├── api_client.ts                       # API methods
└── stores.js                           # State management

API Endpoints

Endpoint	Method	Description
`/voice-calibration/agents`	GET	List available agents
`/voice-calibration/tournament/start`	POST	Start tournament
`/voice-calibration/tournament/{id}/status`	GET	Poll status
`/voice-calibration/tournament/{id}/variants`	GET	Get variants
`/voice-calibration/tournament/{id}/select`	POST	Select winner
`/voice-calibration/generate-bundle/{project_id}`	POST	Generate bundle
`/voice-calibration/{project_id}`	GET	Get existing calibration

The Voice Bundle Files

Voice-Gold-Standard.md

# Voice Reference: Gold Standard

## Winning Sample
[The full text of the winning variant]

## Authentication Tests

### Test 1: Authenticity Check
Question: Is the narrator EXPERIENCING the scene or EXPLAINING it?
Pass if: Reader feels immersed, not lectured

### Test 2: Purpose Check
Question: Does this scene serve its beat function?
Pass if: Clear narrative purpose beyond "this happened"

### Test 3: Fusion Check
Question: Are voice and knowledge seamlessly integrated?
Pass if: Story Bible details feel natural, not info-dumped

Voice-Anti-Pattern-Sheet.md

# Voice Anti-Patterns

## Forbidden Phrases
- "It's important to note that..."
- "Furthermore..." / "Moreover..."
- "In conclusion..."
- [Any AI-sounding formulations]

## Tone Violations
- Excessive hedging ("perhaps", "might", "seems")
- Over-explanation of character motivation
- Breaking POV for convenience

## Project-Specific Bans
[Writer-defined patterns to avoid]

Phase-Evolution-Guide.md

# Voice Phase Evolution

## Act 1: Setup
- Longer sentences, world-building rhythm
- More descriptive, establishing tone
- Character voice emerging

## Act 2A: Rising Action
- Increasing tension in prose
- Shorter paragraphs during conflict
- Internal thoughts more fragmented

## Act 2B: Midpoint & Beyond
- Clipped, urgent prose in crisis
- Callbacks to earlier language
- Voice reflects character growth

## Act 3: Resolution
- Return to opening rhythm (transformed)
- Emotional resonance prioritized
- Earned payoff language

Implementation Status

Component	Status
Backend service	Complete
API endpoints	Complete
VoiceTournamentLauncher	Complete
VoiceVariantGrid	Complete
VoiceComparisonView	Complete
VoiceBundleGenerator	Complete
VoiceVariantSelector	Complete
api_client.ts methods	Complete
stores.js state	Complete
VoiceCalibration.svelte	TO CREATE
ChatSidebar integration	TO DO

Writer’s Journey - Full creative workflow
GraphRAG Conceptual - Knowledge graph philosophy
GraphRAG Implementation - Technical details
Task: Voice Calibration UI - Implementation spec
Task: Narrative Dashboard - Tension/pacing UI

Documentation created: December 5, 2025 Status: Backend complete, Frontend needs master component