Hi HN,


  Hallucinations remain the single biggest bottleneck for moving LLM applications from "cool demo" to "reliable production." Whether it’s a RAG pipeline inventing
  citations or an autonomous agent fabricating data, the lack of a reliable "truth layer" is costing companies billions in trust and market cap.


  I’m sharing HalluciGuard, an open-source middleware layer (AGPLv3) designed to act as a security and reliability buffer between LLM providers and your end-users.

  GitHub: https://github.com/Hermes-Lekkas/HalluciGuard (https://github.com/Hermes-Lekkas/HalluciGuard)


  How it works
  Instead of just hoping the model is right, HalluciGuard intercepts the LLM response and runs it through a multi-signal verification pipeline:


   1. Factual Claim Extraction: It uses lightweight LLMs to break down a response into discrete, verifiable factual claims.
   2. Multi-Signal Scoring: It evaluates each claim using four distinct signals:
       * Self-Consistency: LLM-as-a-judge verification.
       * Linguistic Heuristics: Detecting "uncertainty language" and high-risk patterns.
       * RAG Cross-Reference: Verifying claims directly against your retrieved documents.
       * Web Verification: Optionally pulling real-time snippets from search engines (like Tavily).
   3. Risk Flagging: It returns a comprehensive GuardedResponse with an overall "Trust Score" and flags specific claims as SAFE, MEDIUM, or CRITICAL risk.


  Key Features
   * Provider Agnostic: Native support for OpenAI (GPT-5.x), Anthropic (Claude 4.6), Google (Gemini 3.1), and local models via Ollama.
   * OpenClaw Integration: We’ve built a native interceptor for the OpenClaw agent framework, allowing you to monitor agent actions and thoughts in real-time.
   * Streaming Support: Performs analysis asynchronously so you don't lose the "real-time" feel of streaming responses.
   * Cost-Optimization Cache: We cache verification results locally to reduce your API bills by avoiding redundant checks for common facts.
   * LangChain Ready: Includes a drop-in CallbackHandler for existing LangChain projects.

  Why AGPLv3?
  We believe the "Truth Layer" of the AI stack should be owned by the community, not hidden behind a proprietary corporate API.


  What’s Next?
  We are currently working on "Lookahead Verification" (v0.9), which will attempt to auto-correct hallucinations during the token generation phase before the user
  ever sees them.


  I'd love to get the community's feedback on our scoring heuristics and hear about the edge cases you're seeing in production.