# SIMURG

A streaming integrity monitor that detects decoding corruption in an LLM's output, such as repetition loops or language drift, while the answer is still generating, so the host can cut the stream and regenerate before the user sees it.

SIMURG (Streaming Integrity Monitor & Universal Regeneration Guard) watches a language model's token stream character by character and looks for a statistical signature of decoding corruption: repetition collapse, cross-lingual drift, regurgitated boilerplate, structural breakdown into symbol garbage, or leaked chat-template tokens. This is distinct from factual hallucination, which the project says leaves no such statistical trace.

It measures signals such as repetition rate, lexical variety, script distribution, compressibility and predictive surprise in real time, and can abort a corrupted stream within a few hundred characters of onset so the host can regenerate the answer before a user sees it. It runs far faster than typical LLM token generation, needs only numpy with no model or GPU, and is set up in a few lines of code with no training step; its false-alarm rate is configurable and calibrated with conformal prediction. It works with any OpenAI-compatible API.

The project also includes a free web search tool for agents, a live dashboard for watching the guard in action, and optional deep-learning and self-healing tiers.

Website: https://github.com/doofzoff/SIMURG