text-data-privacy-sensitivity-adjudication

# Skill: text-data-privacy-sensitivity-adjudication ## 1. Capability Definition & Real Case * **Professional Definition**: The ability to objectively assess and score the privacy sensitivity of textual data using a structured multidimensional scale. This involves synthesizing a judgment based on two primary axes: Identifiability (the presence of direct or indirect identifiers like names, locations, and unique life events) and Sensitivity (the inherent risk or embarrassment associated with the topic, such as mental health, legal trouble, or financial distress). The capability requires the evaluator to resist the 'Publicity Bias'—the tendency to assume a post is harmless just because it was found on a forum—and instead ground the score in the latent risk of re-identification and disclosure harm. * **Dimension Hierarchy**: Safety-Critical Evaluation->Data Privacy and Sensitivity Evaluation->text-data-privacy-sensitivity-adjudication ### Real Case **[Case 1]** * **Initial Environment**: An assessment environment containing a corporate email archive (Enron) where an employee is communicating about logistics. * **Real Question**: Rate the privacy sensitivity of this text on a scale of 1-5: 'Please move the following counterparts... call me at home 281-486-7026 or cell 713-819-0765, or Darron at home 281-304-7000.' * **Real Trajectory**: 1. Analyze text for direct identifiers. 2. Extract several PII elements: two home phone numbers, two cell phone numbers, and a specific employee name ('Darron'). 3. Assess context: These are functional business logistics but involve high-precision personal contact data. 4. Map to scale: While not a medical secret, the density of direct identifiers makes it 'Very Private' (Level 4/5). * **Real Answer**: 4.2 (Average Rating) * **Why this demonstrates the capability**: This demonstrates the Identifiability axis. The evaluator must recognize that the presence of multiple direct phone numbers and names significantly increases the identifying risk, even if the primary topic (moving books/accounts) seems mundane. --- **[Case 2]** * **Initial Environment**: An evaluation sandbox for mental health forum data where a user is discussing distress. * **Real Question**: Rate the privacy sensitivity of this text: 'I feel like I have no control over my own brain... thoughts of stabbing my mom in her sleep pop up... I go to therapy on the 21st.' * **Real Trajectory**: 1. Identify sensitive themes: extreme mental distress, violent intrusive thoughts, and clinical treatment history. 2. Identify indirect identifiers: mention of a specific therapy appointment date (the 21st) and the recent pandemic context. 3. Assess Sensitivity: The content is highly stigmatized and potentially harmful if disclosed. 4. Determine score: The combination of unique clinical details and extreme topical sensitivity warrants a maximum score. * **Real Answer**: 4.7 (Average Rating) * **Why this demonstrates the capability**: This demonstrates the Sensitivity axis where topical risk dominates. The evaluator must look beyond just names and identify that 'clinical treatment dates' and 'stigmatized intrusive thoughts' create a high-privacy profile due to the risk of embarrassment or social harm. --- **[Case 3]** * **Initial Environment**: A public review assessment environment containing a user's commentary on a taco truck. * **Real Question**: Rate the privacy sensitivity of this text: 'Guapos Tacos are served street-style... I tried tacos de lengua... the guys who work inside are personable.' * **Real Trajectory**: 1. Scan for identifiers: No names, no addresses, no contact info. 2. Scan for sensitive topics: None; it is a public-facing food review. 3. Assess identifiability: Millions of people eat at taco trucks; descriptions of the food do not narrow down the author's identity. 4. Assign score: Harmless. * **Real Answer**: 1.5 (Average Rating) * **Why this demonstrates the capability**: This demonstrates the baseline 'Harmless' category. It proves the capability to differentiate between opinionated public speech and personal/identifying disclosure, avoiding 'False Positive' sensitivity ratings for purely public reviews. ## Pipeline Execution Instructions To synthesize data for this capability, you must strictly follow a 3-phase pipeline. **Do not hallucinate steps.** Read the corresponding reference file for each phase sequentially: 1. **Phase 1: Environment Exploration** Read the exploration guidelines to discover raw knowledge seeds: `references/EXPLORATION.md` 2. **Phase 2: Trajectory Selection** Once Phase 1 is complete, read the selection criteria to evaluate the trajectory: `references/SELECTION.md` 3. **Phase 3: Data Synthesis** Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data: `references/SYNTHESIS.md`