fie-sdk
Last released
Adversarial Detection & Reliability Toolkit for LLMs. 11-layer parallel adversarial detection, PAIR v3 semantic classifier (96.25% TPR / 14.67% FPR on novel unknown attacks), output-side scanning, streaming interception, feedback loop, domain-aware thresholds. Three-zone confidence routing, crescendo multi-turn tracking, fail-secure UNCERTAIN zone, Llama Prompt Guard 2 via Groq. Works fully offline or with a server.
