32% hallucination rate on a research tool people use to make trades. That was Research v1.5. @SurfAI just shipped Research 2.0 on May 19 and brought it down to 8%. The number that actually matters more is report logic consistency: 55% to 94%. In v1.5, more than 4 in 10 outputs contradicted themselves somewhere in the same report. That is not a minor quality issue. That is a product that actively misleads users who read the whole thing. The rebuild fixed the orchestration layer from scratch, not just the model. What data gets pulled, how many sources get cross-referenced, how the answer gets structured. 37 new API endpoints added across prediction markets, institutional capital flows, DeFi comparisons, and Hyperliquid coverage going back to genesis. The Polymarket and Kalshi integration is the one I find genuinely interesting. v1.5 could not answer a single prediction market question. v2.0 can now scan 30 days of 5-minute BTC markets and flag probability reversals in the final 60 seconds before settlement. That is a specific, actionable edge that did not exist two weeks ago. What I want to see next is whether the 8% hallucination rate holds when the portfolio personalization layer ships. Connecting user positions and trading style to the research agent adds a lot of surface area for the model to go wrong in ways that matter more than generic errors.