Trust centre

• Filtering and monitoring

  Met

  Products used by pupils must not undermine the school's filtering and monitoring duties.

  Evidence: Safe AI complements Smoothwall and Senso rather than replacing them; monitoring alerts route to the named DSL, never to leadership generically.

• Cyber security

  Partial

  Protect all devices and accounts with multi-factor authentication where available.

  Evidence: Role-based access and audited identity reveals are in place.

  Gap: MFA depends on the school's identity provider; SSO (REQ-147) is an interface only.

• Cyber security

  Met

  Maintain an accurate record of what data is held and where it is processed.

  Evidence: The residency register publishes every data category, its store and its processing region.

• Cyber security

  Not yet

  Hold a recognised cyber security certification appropriate to the service.

  Evidence: Certification path published: Cyber Essentials Plus, then ISO 27001, then SOC 2.

  Gap: REQ-174: no certification held today. Publishing the gap is the point of this page.

• Generative AI in education (January 2026) · new January 2026

  Met

  Schools must know which model providers process pupil data and be told when that changes.

  Evidence: The model routing table names every declared route, its provider and whether it is connected.

• Generative AI in education (January 2026) · new January 2026

  Met

  Pupil data must not be used to train foundation models without explicit, informed consent.

  Evidence: Training on pupil data is off by default and cannot be enabled without recorded consent; pupil IP is never used commercially.

• Generative AI in education (January 2026) · new January 2026

  Met

  AI-generated content presented to pupils must be subject to human review appropriate to its risk.

  Evidence: No atom reaches a pupil without specialist review; marking stays teacher-facing until departmental agreement clears the threshold.

• Generative AI in education (January 2026) · new January 2026

  Not yet

  Safeguarding classifiers must have published performance characteristics.

  Evidence: An independent audit with precision and recall by risk category is committed to.

  Gap: REQ-152, blocked by DEC-003. No audited figures exist yet and none are claimed.

• Digital accessibility

  Met

  Services must meet WCAG 2.2 AA.

  Evidence: AA contrast, keyboard paths, reduced-motion support and age-adaptive rendering throughout.

• Data protection

  Met

  A DPIA must be completed before processing children's data at scale.

  Evidence: A pre-filled DPIA is supplied to the school rather than left for the school to write.

• 1. Best interests of the child — Every routing rule is written so the pupil is the beneficiary, not the buyer.
• 2. Data protection impact assessments — A pre-filled DPIA is supplied to every school.
• 3. Age-appropriate application — Rendering, pacing and extension content adapt by year group; extension is suppressed below KS3.
• 4. Transparency — The pupil is told in their own reading level what the tutor records and who can see it.
• 5. Detrimental use of data — No data is used against the pupil's interests; leadership sees aggregates only.
• 6. Policies and community standards — Published policies are the ones actually enforced in the tutor policy engine.
• 7. Default settings — High privacy by default; training off by default; pseudonymity in the corpus.
• 8. Data minimisation — Cache keys exclude pupil identity; identity reveals are audited events.
• 9. Data sharing — Sharing is limited to the school and statutory safeguarding routing.
• 10. Geolocation — Not collected.
• 11. Parental controls — Where a school enables parental visibility, the pupil is told.
• 12. Profiling — Profiling is used for teaching diagnosis only and never for commercial targeting.
• 13. Nudge techniques — No streaks, no loss aversion, no parasocial framing; sessions have natural stopping points.
• 14. Connected toys and devices — Not applicable.
• 15. Online tools — Pupils can exercise their rights through the school; the tutor explains how.

Controller and processor

• Teaching, tutoring and mastery records — controller: The school (or trust); processor: Safe AI. Safe AI acts only on the school's documented instructions.
• Safeguarding records and referrals — controller: The school (or trust); processor: Safe AI. The DSL owns the record. Safe AI routes and retains it; it never decides the outcome.
• Staff accounts and billing — controller: Safe AI; processor: —. Safe AI is controller for its own contractual and account data.
• Operational telemetry with no pupil content — controller: Safe AI; processor: —. Service reliability data only; no pupil content and no pupil identifiers.

1. Nature of the processing

   Safe AI processes pupil work, tutor dialogue and derived mastery records to provide Socratic tutoring, and processes safeguarding classifications to route concerns to the school's Designated Safeguarding Lead. Processing is carried out on the school's instructions.

2. Scope and data categories

   Categories: pupil work, tutor transcripts, mastery records, safeguarding records, staff account data, operational telemetry. Special category data may arise within safeguarding disclosures.

   For the school to complete: Year groups and subjects in scope, and the expected pupil headcount.

3. Context and expectations of children

   Pupils are aged 4–16. The ICO Children's Code is applied in full: no nudge techniques, no profiling for commercial purposes, high privacy by default, age-appropriate explanation of what the tutor records.

4. Purposes and lawful basis

   Public task (education) for teaching and learning; legal obligation and substantial public interest (safeguarding) for concern routing. Consent is not relied on for core teaching, so a pupil cannot be disadvantaged by withholding it.

   For the school to complete: Confirm the lawful basis recorded in your own ROPA.

5. Necessity and proportionality

   Data minimisation: renderings are cached without pupil identity; leadership sees aggregated statistics only; identity reveals (for example book-scrutiny export) are logged as audited events.

6. Risks identified

   Over-reliance on the tutor; misclassification of a safeguarding concern; unauthorised identity reveal; residency gap on classification while REQ-141 is blocked; incorrect content reaching a pupil.

7. Measures to reduce risk

   Anti-dependency reporting to the teacher; four-tier safeguarding routing with DSL ownership; immutable audit log; deterministic verification with a second reader that can only downgrade; specialist review gate before any atom reaches a pupil.

8. Retention and deletion

   Tutor transcripts and mastery records for the duration of enrolment plus one academic year. Safeguarding records to the pupil's 25th birthday, in line with DfE guidance.

   For the school to complete: Confirm against your own retention schedule and record any local variation.

9. Sign-off

   Prepared by Safe AI for completion by the school.

   For the school to complete: DPO name, date, residual risk accepted, review date.

• Item provenance and board-material licensing

  Blocked by DEC-002

  Every item carries provenance; board materials are never redistributed without a licence, and licensing conversations start in year one.

  Next: Open licensing conversations with AQA, Edexcel and OCR before any redistribution.

• Self-hosted safeguarding classifiers

  Blocked by DEC-003

  High-volume safeguarding classification on self-hosted small models, where latency, cost and UK residency all matter.

  Next: Until this ships, the classification route runs on a hosted model and the residency gap is published.

• MIS integration — Wonde first

  Interface only

  Wonde first, covering Arbor, Bromcom and SIMS; Groupcall Xporter as the alternative.

  Next: Wonde application registration and a pilot trust's approval.

• Entra ID and Google Workspace SSO

  Interface only

  Identity through Microsoft Entra ID and Google Workspace, while the MIS stays authoritative for classes and enrolment.

  Next: App registration in both directories; enrolment continues to come from the MIS, never from the directory.

• Independent classifier audit

  Blocked by DEC-003

  An independent audit of the safeguarding classifiers with precision and recall published by risk category.

  Next: Commission the audit once the classifier stack is fixed; publish the figures whatever they say.

• Certification path

  Interface only

  Cyber Essentials Plus, then ISO 27001, then SOC 2 when the US motion starts.

  Next: Cyber Essentials Plus assessment booked before the first trust-wide deployment.