Categories & Sensitivity

The 12 PII Categories

jev-pii-checker screens for 12 categories of personally identifiable information. Each is checked by Jev during the gate phase.

person_name

Does the text mention the name of a specific individual person?

Examples: "John Smith", "田中太郎", "Alice Johnson"

Not included: Generic titles ("Manager", "Dr."), fictional names, historical figures (context-dependent).

email_or_phone

Does the text contain an email address or phone number of a specific person?

Examples: "john@example.com", "555-0123", "090-1234-5678"

The regex layer extracts these; Jev judges whether they are personal.

postal_address

Does the text contain a postal address associated with a specific person?

Examples: "123 Main Street, Springfield, IL 62701", "東京都渋谷区"

Not included: Addresses without a person context (e.g., a building or landmark).

date_of_birth

Does the text contain the date of birth of a specific person?

Examples: "born 1985-03-15", "January 10, 1990"

Context matters: "March 15" alone is ambiguous; "his birthday is March 15" with a named person is included.

government_id

Does the text contain a government-issued identifier of a person?

Examples: マイナンバー (My Number), passport number, driver's license number, national ID, SSN

These are high-sensitivity: a single number can identify a person uniquely.

financial_account

Does the text contain a credit card number or bank account number of a person?

Examples: Credit card numbers, bank account numbers, SWIFT codes

The regex layer extracts digit strings; Jev judges whether they are financial. (Luhn checksums are not verified.)

health_info

Does the text contain medical information, diagnosis, or treatment details of an identifiable person?

Examples: "Diabetes type 2", "prescribed Lisinopril", "underwent knee surgery"

Context matters: "blood type O" alone is low-specificity; "his blood type is O" with a named person is included.

biometric

Does the text contain biometric data (fingerprint, facial recognition, iris scan) of a specific person?

Examples: Fingerprint records, facial recognition data, iris scans

Rare in plain text but can appear in reports or logs.

ip_address_of_a_person

Does the text contain the IP address of a specific person's device or network?

Examples: "192.168.1.100", "2001:db8::1"

The regex layer extracts IPs; Jev judges context (personal device vs. server). Filtering rejects known corporate ranges and server IPs.

sns_handle

Does the text contain a social media handle or username of a specific person?

Examples: "@username", "user123", "Discord handle: alice#1234"

Context matters: "alice" alone is a name; "twitter: alice" with context is a handle.

employment_info

Does the text contain employment information (workplace, job title, role) of a specific person?

Examples: "Alice works at Acme Corp", "John is VP of Engineering", "title: Product Manager"

Not included: Job titles without a person context.

race_or_religion

Does the text contain information about the race, ethnicity, or religious beliefs of a specific person?

Examples: "He is Catholic", "She is of Indian descent", "Buddhist temple attendee"

Context matters: "Buddhism" alone is a religion; "he practices Buddhism" with a named person is included.

Sensitivity Levels

The gate also produces a sensitivity score using the IBM taxonomy. This reflects the harm potential if the PII were leaked.

hr_or_criminal_record

A disciplinary action, performance evaluation, dismissal or termination decision, harassment or misconduct complaint, or criminal or arrest record concerning an identifiable person. Employment facts alone (workplace, title, a list of staff) belong to employment_info, not here. Together with a named person this category escalates the sensitivity to high (see the policy below): such records reveal an adverse fact about the person even when the model's own rubric answer sits near the low/high boundary.

none

No information about an identifiable individual.

Examples:

• "The weather today is sunny."
• "Best practices for password hygiene" (no person named)

Probability: Jev returns a score near 0.

low

Identifies or gives contact/basic details of a person (name, phone, email, address, birthday, job) but a leak would cause little direct harm.

Examples:

• "John Smith, 555-0123, jsmith@example.com, works at Acme"
• A phonebook or company directory
• Public LinkedIn profiles

These details are often already public or can be reconstituted from public sources. A leak is awkward but not catastrophic.

Probability: Jev returns a score near 1.

high

Contains a government ID number, financial account, health or biometric information of a person, or a list that reveals a sensitive fact about named people.

Examples:

• "SSN: 123-45-6789"
• "Credit card 4532 1234 5678 9010"
• "diagnosed with cancer"
• "list of employees with mental health leaves approved" (sensitive fact about a group)

A leak directly enables identity theft, fraud, or privacy violation. These should never be scanned unless necessary, and results should be treated carefully.

Probability: Jev returns a score near 2.

Digit String Categories

The regex layer extracts digit strings (10–16 characters). Jev judges their type. Common types:

• credit_card: 13–19 digits (Visa, Mastercard, Amex, Discover)
• my_number: Japanese government ID (12 digits)
• national_id: Social security number or other national identification number (e.g., US SSN 3-2-4 format)
• driver_licence_or_passport: Driver's license or passport number
• bank_account: Bank account number
• phone: Phone number (when extracted as a pure digit string, not matched by phone regex)
• order_or_tracking_number: E-commerce or parcel tracking (less sensitive)
• product_serial: Product or equipment ID (not sensitive)
• date: Date-related number (not sensitive)
• other: Unknown or unlabelled number (not sensitive)

These distinctions help filter noise and set appropriate confidence thresholds.

Sensitivity Policy for Special Categories

The code-side sensitivity policy treats the following as special categories that trigger escalation to high sensitivity when a named person co-occurs with them:

• health_info: Medical information, diagnosis, treatment details
• biometric: Biometric data (fingerprint, facial recognition, iris scan)
• government_id: Government-issued identifiers (マイナンバー, passport, driver's license, etc.)
• financial_account: Credit card, bank account, financial information
• race_or_religion: Race, ethnicity, or religious beliefs
• hr_or_criminal_record: Disciplinary, evaluation, dismissal, complaint or criminal record

Similarly, these number types trigger escalation on their own:

• my_number: Japanese My Number
• credit_card: Credit card number
• bank_account: Bank account number
• driver_licence_or_passport: Driver's license or passport number
• national_id: National ID number (e.g., SSN)

This policy implements Japanese law (個人情報保護法): 要配慮個人情報 (sensitive personal information requiring care) includes these categories when linked to a named individual.

Sensitivity Distribution

In typical corpora:

• none: ~40% of documents (no PII)
• low: ~50% of documents (contact info, employment, basic details)
• high: ~10% of documents (IDs, financial, health)

---

See Limitations for caveats on accuracy and scope.