LLM Hallucination & Reliability Simulator

Large language models generate text one token at a time, sampling from a probability distribution rather than looking facts up. This simulator visualizes that process as a path through a 3D "knowledge space": a cluster of glowing points represents verified facts a retrieval system could supply, and the traveling particle is the model mid-generation. Turn up temperature and the path wanders freely — sometimes straight into unsupported claims, shown in red. Turn up RAG grounding strength and the path clings to the nearest verified fact, shown in blue. Toggle fact-checking to watch a Constitutional-AI-style correction pass pull hallucinated segments back toward the evidence — the same mechanism behind RLHF fine-tuning and automated fact-checking pipelines used to make real LLMs more reliable.