Neuro-Symbolic Architecture

The Two Layers

Neural Layer (LLM)

• Handles narrative, dialogue, improvisation
• Parses player intent into structured actions
• Wraps mechanical results in story

Symbolic Layer (Python)

• Enforces SRD 5e rules strictly
• Uses RNG for dice (never LLM prediction)
• Validates all state changes

The Resolution Loop

Player Input ("I attack the goblin")
    ↓
Neural: Parse Intent → AttackAction(target="Goblin")
    ↓
Symbolic: Execute Rules (roll dice, check AC, calculate damage)
    ↓
Result: {success: true, damage: 8}
    ↓
Neural: Narrate → "Your blade catches the goblin's shoulder..."

Implementation Patterns

Skill Functions (Symbolic)

def resolve_attack(attacker: Entity, target: Entity) -> AttackResult:
    """Pure function. Takes data, returns data. No LLM calls."""
    roll = roll_dice("1d20")
    # ... SRD logic ...
    return AttackResult(hit=True, damage=8)

Agent Prompts (Neural)

PARSE_INTENT_PROMPT = """
Given the player's action, extract the structured intent.
Output JSON matching the ActionIntent schema.
"""

NARRATE_RESULT_PROMPT = """
Given the mechanical result, write engaging narrative.
Maintain the tone and pacing of the scene.
"""

Core Axiom

Dolt is for Truth, Neo4j is for Search

• Kill a goblin → Event in Dolt (can rollback/fork)
• Goblin's brother hates you → Relationship in Neo4j (for retrieval)