## Turn 1

Synthetic benchmark scenarios inspired by email triage.
No real private data.
Numbered list of 5 scenarios.
Benchmark harness (internal context, but doesn't change the task much, just means I should be precise and high-quality).

    *   What does email triage involve?
        *   Classification (Spam vs. Ham, Urgent vs. Low Priority).
        *   Sentiment analysis (Angry customer vs. Inquiry).
        *   Entity extraction (Dates, Order IDs, Names).
        *   Summarization (Long threads into a few bullet points).
        *   Action item extraction (What does the user need to *do*?).
        *   Language translation/detection.
        *   Routing (Which department should this go to?).

    *   *Scenario 1: Classification and Priority.* (Standard triage).
    *   *Scenario 2: Sentiment and Escalation.* (Customer support focus).
    *   *Scenario 3: Action Item Extraction.* (Project management focus).
    *   *Scenario 4: Summarization of Long Threads.* (Executive summary focus).
    *   *Scenario 5: Entity Extraction and Routing.* (Logistics/Data entry focus).

    *   *Scenario 1 (Classification):* A mix of marketing, newsletters, personal, and spam. Goal: Categorize and flag "Action Required."
    *   *Scenario 2 (Sentiment):* Customer complaints about a late shipment. Goal: Identify the emotion and determine if it needs immediate human intervention.
    *   *Scenario 3 (Action Items):* A project kickoff email with multiple tasks. Goal: Extract a checklist of tasks and deadlines.
    *   *Scenario 4 (Thread Summarization):* A 10-email long technical bug report thread. Goal: Summarize the current status and the "blocker."
    *   *Scenario 5 (Routing/Entity Extraction):* Invoices and purchase orders. Goal: Extract Order ID, Amount, and Vendor, then route to the correct department.

    *   Are there any real names/emails? No, all synthetic.
    *   Is it a numbered list? Yes.
    *   Are there 5? Yes.