🏆 Inference Collapse: Real-Time Hallucination Audit

🏆 Inference Collapse: Real-Time Hallucination Audit

Inference Collapse is a cyberpunk-inspired real-time simulation game that binds Gemma 4 (31B-it) inference outputs directly to downstream world-state behavior.
The project visualizes LLM hallucination, overconfidence, and unstable reasoning as physical threats, spatial corruption, and environmental instability inside an interactive game system.

💡 Project Overview

In this game, the player acts as an audit bot exploring corrupted sectors while monitoring inference signals generated by Gemma 4.

As reasoning quality degrades, the game world itself becomes unstable.

Structured inference signals from Gemma 4 are directly injected into gameplay logic:

confidence → enemy AI speed, field of view (FOV), and player visibility changes
severity → spatial glitches, coordinate noise, and world instability

This creates a real-time relationship between:

LLM inference → downstream world-state corruption

Rather than treating hallucination as a simple model failure, this project turns it into an interactive gameplay mechanic.

🧠 Why Gemma 4 (31B-it)?

This project intentionally uses Gemma 4 – 31B-it (Dense) because of three key advantages:

1. Strong reasoning consistency

To simulate believable hallucinated logic, the model needs to maintain coherent reasoning while still producing biased or unstable conclusions.
31B-it provided a strong balance between reasoning depth and output stability.

1. Structured JSON outputs

The game directly binds inference values such as confidence and severity into real-time physics and AI behavior.
Reliable structured output was essential for stable simulation updates.

1. Future local deployment potential

Although the current version runs through API-based inference, the architecture was designed with future local Gemma deployment in mind.

🛠 Technical Implementation
Backend
Python
Streamlit
Google GenAI SDK

Gemma 4 generates structured inference signals that are parsed and transformed into gameplay parameters.

Frontend
HTML5 Canvas
JavaScript

Using a real-time rendering loop, Gemma’s outputs are directly mapped into movement logic, visibility systems, and environmental distortion.

Inference Pipeline

Gemma output → JSON parsing → world-state binding → gameplay update

Fallback Mode

If API access becomes unavailable (degraded mode), the system switches to a local simulation fallback so reviewers can still experience the inference-driven mechanics.

🎨 Originality

This project redefines hallucination not as a passive AI error, but as:

Hallucination as a Gameplay Mechanic

False confidence, unstable reasoning, and contradictory outputs directly affect enemy behavior, visibility, and environmental corruption.

The goal was to create an experience where LLM reasoning quality itself becomes the difficulty system of the game.

GitHub: Inference Collapse