Detect moderation-risk signals using toxicity patterns, coordinated bursts, policy triggers, and creator context.