Eval frameworks that work in research don't always survive contact with production financial data. This session brings together panelists covering what they measure, what they have stopped measuring, how they handle non-determinism at scale, and what broke before they got it right. What this session will cover: - Why standard benchmarks fail on production financial datasets - How teams are handling non-determinism in multi-step agent pipelines - Observability tooling being used in live financial environments - What good eval hygiene looks like six months into production Bring your own war stories, this one tends to turn into a comparison of notes.