Public transparency artefact
EduZen AI Tutor — AI System Card
Version 0.1.0-foundation · Issued 7/8/2026, 9:24:01 AM
Purpose
Adaptive tutoring for GCSE / 11+ / KS3 learners aged 9-18. The system explains, hints and assesses but does not award formal grades.
Risk class: high-risk (EU AI Act Annex III §3 — education)
Model routing
- default tutorclaude-sonnet-4.5
- mark feedbackgpt-5.2
- image ocrgemini-3-flash
frontier large language models accessed via the Emergent Universal Key — Anthropic Claude, OpenAI GPT, Google Gemini
We do not train models on children's data. Period.
Guardrails
- PII pre-filter and post-filter (regex + classifier + LLM-judge)
- Refusal taxonomy: out_of_scope, unsafe, ambiguous, off_curriculum, requires_human, requires_parent, regulated_advice, exam_cheat
- Confidence score (logprobs + self-consistency n=3) — answers below 0.6 are flagged
- Hallucination tripwire against curated curriculum knowledge base
- No cross-session memory for under-13 by default
Human oversight
- Teacher-in-the-loop for school deployments (transcript visibility)
- Parent-in-the-loop for B2C (weekly digest + on-demand transcripts)
- Monthly red-team panel + safeguarding review
Safeguarding
Safeguarding overlay shown to child; session frozen; Designated Safeguarding Lead alerted within 5 minutes.
Audit logging
Retention: 7 years · Storage: MongoDB append-only with hash chain (sha256, prev_hash linked).
Show logged fields
log_id · ts · user_id_hash · session_id · role · child_age_band · prompt_hash · prompt_category · model · model_version · system_prompt_hash · response_hash · confidence · refusal_flag · safety_flags · latency_ms · cost_tokensContact
- dpo emaildpo@eduzen.ai
- safeguarding emailsafeguarding@eduzen.ai
- support emailsupport@eduzen.ai