Internal tool, not indexed. Same code the API runs before every model call.

PII firewall tester

Type anything a teacher might. Left: what you typed, with detected personal information marked. Right: exactly what would be sent to the AI model. Nothing here leaves your browser.

Teacher types

Write a report comment for Aroha (she/her), Year 4. She reads confidently, struggles with fractions and helps Ben when he's stuck. Mum is on aroha.mum@gmail.com. Next term we study Kate Sheppard.

Model receives

Write a report comment for Student_1 (she/her), Year 4. She reads confidently, struggles with fractions and helps Student_2 when he's stuck. Mum is on [email_1]. Next term we study Kate Sheppard.

Student_1 ← Aroha
Student_2 ← Ben
[email_1] ← aroha.mum@gmail.com

Rules fired: strong-trigger, ambiguous-person, email. The map above never leaves the server; the reply is swapped back before the teacher sees it.

Known names (from structured fields, always swapped)

Allowed names (never swapped)

How it decides

Names come from a 22,000-entry first-name list (with Māori and Pasifika names added), context words like "student" or "report comment for", class-list shapes, possessives, and the names you give it. Names that are also everyday words (Grace, Hope, Summer) need context. Full names outside a student context are left alone so the model still knows who Kate Sheppard is. Emails, phones, student numbers, dates of birth and addresses are matched by shape.