I test whether your RAG or retrieval-backed AI system actually stays grounded in its supplied evidence. I probe for unsupported claims, fabricated citations, citation/source mismatch, source authority confusion, missing uncertainty, retrieval failures, contradictory evidence handling, and answers that exceed the available evidence. You receive an adversarial grounding test set, a pass/fail rubric, documented failure examples, source and citation integrity findings, and reusable regression cases — delivered as structured JSONL/YAML plus a human-readable report where technically appropriate. One fixed-price project. I report what I can prove from the evidence, without exaggerated claims.