I've built a Rust-based tool that visualizes token-level confidence and hallucination likelihood from large language model (LLM) outputs.

The goal is to move beyond simple binary classification ("is this hallucinated?") and give users a detailed, interpretable view of where and how an LLM's response may go off the rails.

Key Features:

Token-level confidence visualization (color-coded)

Flagged issues: hallucinations, overconfidence, uncertainty

Terminal rendering (with color)

HTML report generator (interactive)

Markdown output

Stats: confidence range, flagged spans, density

Demo data with realistic hallucination cases

Modular: add custom renderers, integrate into pipelines

Run it via CLI, or integrate as a library into your own tools or dashboards.

GitHub: https://github.com/Mattbusel/LLM-Hallucination-Detection-Scr...

Would love feedback, ideas for integration (e.g. Jupyter, browser extension, VSCode plugin), and thoughts on how to best measure and surface LLM epistemic uncertainty.