Findings

Our own results. Each finding ships its methods in full and links the harness that reproduces it.

• 27 Jul 2026We put Google's search on 480,000 German court decisions and measured itGoogle CSE and Vertex AI Search benchmarked against our retrieval stack on real user queries, 2,941 held-out case-law questions, and 116 landmark research questions. We win two of three, and the loss taught us the most.