Customer sentiment over the last N days (Reports -> Sentiment), one conversation at a time. summary: conversations read (unread = could not be read, left out of rates), ending_negative_rate, net_sentiment (% closing positive minus % closing negative, -100..100), opened_negative and recovery_rate (of the conversations that opened negative, how many closed neutral or positive), worsened_rate, attention_now. matrix: opening -> closing counts. breakdown.agents credits the outcome to whoever handled the conversation at its last burst (handler bot = the AI agent); judge agents by recovery_rate, never by how many negatives they inherited, and say when a row is low_sample. Also frustration by inbox, flags (asked for a person, churn risk, abusive language, minutes until an agent replied), quality (CSAT by closing sentiment, response and resolution times by opening sentiment) and the conversations that need attention now (attention: the 10 most recently active; summary.attention_now is the total). Sentiment is recorded only when it runs in enforce; all zeros usually means it is off or still in shadow -- say so. Call it "sentimiento" in Spanish, never "ánimo" (frustration levels: tranquilo, molesto, frustrado, hostil). Suggest reviewing the AI agent only when breakdown.agents has a bot row with conversations; a conversation an agent handled or nobody did needs a reply or an assignee, not a bot check.