Where Logic
Meets Magic.

NarrativeNode: The Systemic RPG Engine

The primary bottleneck in Generative AI gaming is consistency. Standard LLM-based games suffer from "hallucinations"—where the AI forgets established lore, breaks game rules, or generates illogical outcomes in favor of creative writing.

💡 Our Solution: The "Judge-First" Architecture

NarrativeNode implements a proprietary multi-agent pipeline designed to enforce logic within an open-ended world. Unlike traditional systems that rely solely on a creative writer agent, we separate Intent, Logic, and Narration.

1. The "AI Judge" System (Intent Analysis)

Instead of letting a massive LLM hallucinate a response immediately, the player's input first passes through a lightweight Judge Agent.

• Function: The Judge analyzes the raw player input to extract intent and actions (e.g., "Attack the guard" vs. "Persuade the guard").
• Logic (The Python Middleware): The Judge's structured output is sent to the Python Engine. The engine calculates the outcome based on game stats, RNG, and world rules—completely deterministic and hallucination-free.
• Result: The Engine constructs a specific, optimized prompt for the Narrator. This ensures the story always follows the game's mathematical rules and significantly reduces token costs by filtering unnecessary context.

2. Iterative Fine-Tuning Pipeline

• Models are trained specifically on our custom datasets to understand not just "how to write," but "how to adhere to game mechanics."
• Every development phase contributes new data to refine the model's adherence to the project's unique tone (Dark Fantasy / Lovecraftian).

3. Constrained Player-Driven Generation

• Dynamic Assets: Whether generating unique items, locations, or lore, the engine applies "Style & Logic Wrappers" to ensure player creations fit the game's tone and balance.
• Rule Compliance: A player can request anything, but the engine only generates what is possible within the simulation's logic.

---

🔧 Tech Stack

• Orchestration: Python-based Multi-Agent System (Judge, Narrator, World Manager).
• Logic Core: Graph Theory for non-Euclidean location tracking (Mapless Navigation).
• AI Integration: Custom Fine-Tuned adapters for Claude 3.5 Sonnet (Creative) and optimized SLMs (Logic).
• Client: Godot Engine (4.x) for UI and Visual Feedback.

🗺 Roadmap

• Phase 1: Proof of Concept: Implementing the "Judge -> Engine -> Narrator" pipeline to filter hallucinations.
• Phase 2: Fine-Tuning: Training smaller models (SLMs) for cost-effective rule enforcement and intent analysis.
• Phase 3: Cloud Infrastructure: Deploying the multi-agent swarm on AWS/GCP for scalable inference.
• Phase 4: Alpha Release: Opening the "Systemic RPG" to early testers.