Use when an A/B test result is read for a ship decision, especially off a dashboard, a small sample or many metrics. Computes significance per metric with multiple-comparison correction, reads confidence intervals, checks novelty, interactions, Simpson's paradox, guardrails and stop-early rules.