Narrative Theory as Code

The core innovation of the [[denote:20251019T134950][drama project]]: converting classical narrative theory into executable validators.

Concept

Instead of hoping generated stories follow dramatic principles, test them programmatically using narrative theory as the specification.

Think of it as test-driven story development where literary theory provides the test suite.

Theory → Code Mappings

Aristotle’s Poetics

Unity of Action:

class AristotelianUnityValidator:
    def validate(self, sketch):
        # Check single coherent action
        # Check appropriate magnitude (not too few/many events)
        # Return pass/fail + feedback

Peripeteia (Reversal):

class PeripeteiaValidator:
    def validate(self, sketch):
        # Check each character arc has reversal of fortune
        # Score = coverage across all arcs

Anagnorisis (Recognition):

class AnagnorisisValidator:
    def validate(self, sketch):
        # Check characters have moment of recognition/discovery
        # Score = coverage across all arcs

Propp’s Morphology

Character Spheres: - Villain, Donor, Helper, Princess, Dispatcher, Hero, False Hero - Validate characters fill required spheres - Check functions (31 narrative functions) appear in sequence

Todorov’s Grammar

Narrative Sequence: - Equilibrium → Disruption → Resolution - Validate proposition sequences - Check temporal coherence

Genette’s Narratology

• Order: Analepsis (flashback), prolepsis (flash-forward)
• Focalization: POV consistency
• Voice: Narrative voice coherence

Implementation

Source material: Todorov’s “Grammaire du Décaméron” (full text in ~/proj/drama/seeds/todorov.txt)

Validators in: ~/proj/drama/src/drama/validators/narrative/

Multiple Test Types

1. Structural - Pydantic schemas enforce Aristotelian rules
2. LLM Judge - Fresh LLM evaluates without generation context
3. Internal - Generating LLM checks its own work
4. Constraint Propagation - Failed tests trigger corrective generation

Granularities

• Micro: Individual propositions (“does this action advance the arc?”)
• Meso: Scene/course level (“does this dinner course contain reversal?”)
• Macro: Complete arcs (“did each guest achieve anagnorisis?”)

Why This Matters

Narrative theory often feels abstract and subjective. By encoding it as tests, we make it: - Concrete: Clear pass/fail criteria - Reproducible: Same story always gets same result - Iterative: Generate → Test → Fix → Re-test

We can also analyze what works: which narrative structures pass tests most reliably? Which LLM strategies produce better dramatic arcs?

Example

sketch = generate_sketch(premise="A mysterious dinner party")

results = [
    AristotelianUnityValidator().validate(sketch),
    PeripeteiaValidator().validate(sketch),
    AnagnorisisValidator().validate(sketch),
]

if all(r.passed for r in results):
    proceed_to_spec_generation(sketch)
else:
    refine_sketch_based_on_feedback(sketch, results)

Related Notes

• [[denote:20251019T134950]] - Drama Project
• Multi-Stage Story Generation with Validation - Multi-Stage Generation

---

Status: Seedling | Created: 2025-10-19