I built an eval-driven financial system that treats “confident wrong answers” as the main failure mode [P]

I built TrustRAG Finance, an evaluation-driven financial research RAG assistant. The goal was not to build another PDF chatbot. I wanted to test a more production-style question: How do you design a RAG system where a confident wrong answer is treated as the main failure mode? Current pipeline: - Hybrid retrieval: BM25 + dense + RRF + rerank. - Provider-neutral LLM synthesis: Gemini / OpenAI / Anthropic / stub behind one interface. - Structured cited answers. - Claim-level citation verification. - Independe