## Turn 1

Synthetic benchmark scenarios inspired by email triage.
No real private data (synthetic/anonymized).
Numbered list of 5 scenarios.
Home-lab benchmark harness (implies a need for complexity, variety, and measurable metrics).

    *   *What are the core tasks in email triage?*
        *   Classification (Spam vs. Ham, Urgent vs. Non-urgent).
        *   Sentiment Analysis (Angry customer, Happy feedback).
        *   Entity Extraction (Names, dates, order numbers).
        *   Summarization (Condensing long threads).
        *   Action Item Extraction (What needs to be done?).
        *   Intent Recognition (Request for refund, technical support, sales inquiry).
        *   Prioritization (Sorting based on SLA or VIP status).

    *   *Scenario 1: Customer Support Classification & Priority.* (Classic triage).
    *   *Scenario 2: Corporate Internal Communication Summarization.* (Long threads, multiple participants).
    *   *Scenario 3: Sales Lead Qualification.* (Identifying intent and urgency from vague inquiries).
    *   *Scenario 4: Security/Phishing Detection.* (Identifying malicious intent or suspicious links).
    *   *Scenario 5: HR/Employee Relations Sentiment & Action.* (Handling sensitive internal requests).

    *   *Scenario 1: Multi-Class Support Ticket Routing.*
        *   Input: A batch of 50 emails from diverse customers.
        *   Task: Classify into categories (Billing, Technical, Shipping, General) and assign priority (P0-P3).
        *   Benchmark Metric: Accuracy, F1-score, and latency per email.

    *   *Scenario 2: Thread Compression and Action Item Extraction.*
        *   Input: A long email thread (10+ replies) regarding a project delay.
        *   Task: Summarize the current status in 3 sentences and list specific action items with owners.
        *   Benchmark Metric: ROUGE score (summary), Extraction accuracy.

    *   *Scenario 3: Sales Lead Scoring & Intent Analysis.*
        *   Input: Inbound inquiries from a "Contact Us" form.
        *   Task: Determine if the lead is "Hot," "Warm," or "Cold" based on budget mentions, timeline, and product interest.
        *   Benchmark Metric: Precision/Recall on "Hot" leads.

    *   *Scenario 4: Phishing and Fraud Detection (Reasoning).*
        *   Input: A series of emails, some containing subtle social engineering tactics.
        *   Task: Identify the "Red Flags" and provide a reasoning for why an email is marked as suspicious.
        *   Benchmark Metric: False Positive Rate, Reasoning Coherence.

    *   *Scenario 5: Executive Briefing & Sentiment Mapping.*
        *   Input: A collection of feedback emails from a recent product launch.
        *   Task: Aggregate the primary complaints, identify the overall sentiment trend, and suggest three immediate improvements.
        *   Benchmark Metric: Sentiment alignment, Number of unique issues identified.

    *   Check against constraints: Synthetic? Yes. No private data? Yes. Numbered list? Yes. 5 scenarios? Yes.