Langfuse shipped a major architectural shift toward observation-centric data modeling while expanding evaluation capabilities. The v4 release rebuilt the core table structure to treat individual operations—LLM calls, tool executions, agent steps—as first-class query units instead of nesting them under traces, unlocking significantly faster filtering and aggregation at scale. This foundation enabled observation-level evaluations that score specific operations rather than entire workflows, and boolean and categorical LLM-as-a-Judge scores that move beyond numeric gradients. Experiments got a speed overhaul with faster UI loading and the ability to run against versioned datasets for reproducibility, while new tools like the CLI and corrected outputs for human-in-the-loop workflows expanded how teams interact with production data.