Capstone: evaluate an LLM feature end to end

Take one LLM feature, build its golden dataset and metrics, check a judge against humans, add a regression gate and a monitoring plan, and report results with uncertainty. Deliverables, rubric and where to go next.