Episode Details

Your retrieval pipeline returns a document, the model cites it, and the claim checks out. Everything in the trace looks fine. But what if the model actually pulled that answer from its training data and the citation is just decoration? This episode digs into a genuine gap in AI observability: tracing shows what your agent did, never why. We walk through our own planning-agent-to-research-agent-to-scriptwriter pipeline as a case study, then dig into new mechanistic interpretability research on how LLMs decide to cite — and why current context-attribution tools can't separate what the model knew from what you handed it. If you're building knowledge-based generative products, this one reframes what your citations actually prove. Episode #874287 — open it directly at myweirdprompts.com/874287