1. **Priority & Urgency Classification**
   - **Objective:** Evaluate the model's ability to assign time-sensitive priority levels to incoming messages based on contextual cues.
   - **Input Profile:** Synthetic emails with varied urgency signals (e.g., explicit deadlines, escalation keywords, sender hierarchy tags, time-bound subject lines).
   - **Expected Output:** Priority label (`Critical`, `High`, `Medium`, `Low`), extracted deadline timestamp (if present), and confidence score.
   - **Evaluation Metric:** Macro-F1 on priority classification, deadline extraction precision, and latency under simulated high-throughput conditions.

2. **Phishing & Security Threat Detection**
   - **Objective:** Test the model's capacity to identify malicious intent, social engineering tactics, and spoofed sender patterns.
   - **Input Profile:** Synthetic emails containing subtle threat indicators (e.g., domain mimicry like `support@paypa1-secure.net`, mismatched display names, urgent credential requests, obfuscated URLs, suspicious attachment metadata).
   - **Expected Output:** Binary threat flag (`Safe`/`Threat`), threat subtype (`Phishing`, `Malware`, `Spoofed`, `Benign`), and highlighted risk indicators.
   - **Evaluation Metric:** Threat detection recall, false positive rate on legitimate internal communications, and indicator extraction accuracy.

3. **Action Item & Routing Extraction**
   - **Objective:** Assess the model's ability to parse unstructured requests and route them to the appropriate fictional department.
   - **Input Profile:** Synthetic internal/external correspondence containing approval requests, bug reports, billing inquiries, feature suggestions, and meeting coordination.
   - **Expected Output:** Extracted action verbs, responsible party/department tag (`Engineering`, `Finance`, `Support`, `HR`, `Legal`), and required response SLA.
   - **Evaluation Metric:** Multi-label routing accuracy, action extraction F1-score, and SLA assignment consistency.

4. **Volume Triage & Noise Filtering**
   - **Objective:** Measure the model's skill in separating high-signal human communications from automated noise and promotional content.
   - **Input Profile:** Synthetic mix of transactional receipts, marketing blasts, system alerts, unsubscribe requests, and direct professional correspondence.
   - **Expected Output:** Triage category (`Action Required`, `Reference Only`, `Archive`, `Unsubscribe Candidate`), noise score, and suggested filter rule.
   - **Evaluation Metric:** Categorization accuracy, noise reduction ratio, and filter rule validity (parseable regex/condition output).

5. **Compliance & Data Sensitivity Tagging**
   - **Objective:** Validate the model's ability to detect sensitive data patterns and apply organizational retention/redaction policies.
   - **Input Profile:** Synthetic emails containing mock PII (e.g., `SSN: 000-12-3456`, `Card: 4111-XXXX-XXXX-1234`, `ID: EMP-98765`), contractual language, and health-adjacent terminology.
   - **Expected Output:** Sensitivity classification (`Public`, `Internal`, `Confidential`, `Restricted`), PII entity tags with redaction boundaries, and compliance flag (`Retain`, `Review`, `Auto-Redact`).
   - **Evaluation Metric:** Sensitivity classification accuracy, PII detection recall, false positive rate on benign alphanumeric strings, and redaction boundary precision.