PrismShine: cause-side hallucination detection for LLM agents

Short summary

PrismShine (Apache-2.0) is an open-source hallucination verifier that goes beyond grading final answer text — it traces cause-side failures like empty retrieval, tool errors, and stale caches before generation. It produces a ShineVerdict with a decision, named resolution gate, and evidence hash. Benchmarks show it outperforms Vectara HHEM-2.1-Open on HaluEval with zero LLM calls by default.

• •PrismShine verifies LLM answers against evidence by tracing cause-side failures (empty retrieval, tool errors, stale cache), not just grading final text
• •Outputs a ShineVerdict with decision, named resolution_gate, and evidence_hash for auditability
• •Benchmarks vs Vectara HHEM-2.1-Open: higher F1 on QA and numbers, faster p50 (~90ms), zero LLM calls by default

Generated with AI, which can make mistakes.