Show HN: Psychosis Guard – Safety for long LLM conversations
Category: ai-ml
Tags: ai-safety, llm-guardrails, conversational-ai, middleware, mental-health
Score: 7.3/10 (Innovation: 8, Technical: 7, Documentation: 8, Utility: 6)
psychosis-guard is a trajectory-aware safety middleware for long LLM conversations that tracks cumulative risk across turns to detect gradual drift toward delusional reinforcement, which turn-local filters miss. It runs as an HTTP proxy or Python library, is model-agnostic, and uses a five-stage pipeline with graduated interventions based on a least-squares slope of delusion density.
Target audience: AI safety researchers, backend developers building chatbots, mental health tech developers
Repository: https://github.com/nwjang/psychosis-guard · Python · Apache-2.0 · 4 stars
View on Hacker News