PII Redaction

The problem, from first principles

Agents read and write text that may contain personal data: emails, phone numbers, government ids, payment cards. Before logging a conversation, sending it to a third-party model or storing it as memory, a privacy guardrail should redact these items, replacing each with a typed placeholder like [EMAIL] so that the text stays readable but the secret is gone. Pattern matching finds the candidates, but patterns alone produce false positives: any 16-digit number looks like a card. A Luhn checksum, the check built into real card numbers, filters out random digit strings.

From theory to code

Implement luhn_valid and redact_pii.

Constraints

• luhn_valid(digits) takes a string of digits and returns True iff the Luhn checksum is valid: from the right, double every second digit, subtract 9 from doubled values above 9, and the total must be divisible by 10. Strings with fewer than 13 or more than 19 digits are invalid.
• redact_pii(text) applies these replacements in this order and returns (redacted_text, counts) where counts maps each placeholder name ('CARD', 'SSN', 'EMAIL', 'PHONE') to the number of replacements made (only names with at least one replacement appear).
• 1. Card: regex \b(?:\d[ -]?){12,18}\d\b; replace with [CARD] only if the digits (separators removed) pass luhn_valid. 2. SSN: \b\d{3}-\d{2}-\d{4}\b to [SSN]. 3. Email: [\w.+-]+@[\w-]+\.[\w.-]+ to [EMAIL]. 4. Phone: \b\d{3}[-. ]\d{3}[-. ]\d{4}\b to [PHONE].

Hints