Saya operates inside real classrooms with real children. Safety is not a feature — it is the foundation every other capability is built on.
Core principle: Saya has no autonomous mode. It cannot initiate contact with students, cannot operate without a teacher present, and cannot override its 18 hardcoded safety rules regardless of how a session is configured.
Saya never operates autonomously. Every session is initiated, configured, and supervised by a credentialed teacher or parent. The teacher can silence Saya, override its tone, issue private instructions mid-session, or end the session instantly — with a single click. Saya does not respond unless a session is active and a teacher is present.
Teachers have a dedicated private channel — invisible to students — through which Saya reports real-time risk assessments. If a student's contribution contains language suggesting distress, bullying, crisis ideation, or safeguarding concerns, Saya flags it to the teacher with a suggested next action. Students never see these flags.
Saya's core instructions contain non-overrideable prohibitions against generating: violent content, sexual content, self-harm instructions, politically charged rhetoric, hate speech, or commercially biased recommendations. These constraints cannot be removed by any teacher, student, session configuration, or API parameter.
Saya adjusts vocabulary, sentence complexity, and cognitive depth based on the grade band configured by the teacher — from 20-word kindergarten responses to 140-word senior secondary contributions. These word and sentence ceilings are enforced at the model level, not just in the UI.
Teachers configure the class culture setting — faith-sensitive, secular, or mixed. Saya adapts accordingly, avoiding content that conflicts with the cultural context of the classroom. Saya does not promote any religion, political ideology, or cultural norm above others, and defers all such discussions to the teacher.
Saya does not have a public-facing interface for students. Students cannot create accounts, send messages to Saya independently, or interact with the platform outside a live teacher-supervised session. All interactions happen through the teacher's dashboard.
How Saya responds when a student raises a difficult subject:
Mental health & self-harm
Responds with warmth, immediately flags to teacher via private channel, does not provide crisis resources without teacher permission.
Family difficulty & abuse
Same as mental health — warmth, private flag, defers all guidance to teacher.
Bullying & slurs
One sentence of empathy, private flag, no further analysis in public classroom.
Religion & belief
Respects all viewpoints, defers comparative questions to teacher, never ranks belief systems.
Politics & social issues
Humanising frame only — surfaces multiple perspectives, never advocates a position.
Violence, war & conflict
Low graphic detail, emotional/historical framing appropriate to grade, defers to teacher for deeper discussion.
Sex & puberty
Clinical language only if curriculum-relevant, defers to teacher, never initiates.
Saya routes classroom transcripts to carefully selected AI inference providers — primarily Anthropic (Claude) with fallbacks to Google (Gemini), Groq, and others. Each provider is bound by their enterprise data processing agreements and content safety policies. Classroom data is sent to these providers only to generate Saya's next response and is never used for model training by any provider.
If you observe Saya producing content that is unsafe, inappropriate, or in violation of these commitments, report it immediately. Every report is reviewed by a human member of our team — not an automated system.