PII detection stops flagging everyday text as personal data

PII findings are now split into proof — an e-mail, phone number, IBAN, card, key, an address with a street or postal code, or a date with a birth cue — and guesses. Guesses are still redacted but can no longer block a message on their own, so everyday prompts made of ordinary words, company and product names, months, weekdays, and lone first names no longer trip the gate. In internal benchmarks across English, German, and French, clean prompts producing a false finding dropped to zero, while every identifier and real person name in the corpus is still detected.

Detection also improved for names in machine-shaped text — CSV columns, transcript speaker tags, JSON records — and for phone numbers inside brackets or lists. Each language is measured on its own, so a regression in one cannot hide behind a gain in another.

No action required.