Safety is not a setting. It is the architecture.
Every prompt and every reply passes through the same five-stage pipeline. Nothing reaches a child unscreened. Nothing leaves the system unlogged.
Pre-screening
Every prompt is classified for intent, sentiment and risk before it reaches a model.
Policy routing
Blocked categories are refused inline; monitored patterns are tagged for review.
Agent response
Curriculum-aligned reply generated within the agent's character and reading age.
Post-screening
The reply is re-checked for safety, tone and accuracy before the child sees it.
Audit & escalation
Severity-scored, logged, and escalated to a human safeguarding lead when needed.
What children and parents can see.
5-stage pipeline
Pre-screen, route, respond, post-screen, escalate.
Childline fallback
Always-on signposting to 0800 1111.
Per-agent redirects
Each character holds the same calm, scripted safe responses.
What runs behind every reply.
Pre/post-screening
Every message is checked before and after generation.
Severity scoring
Signals are scored 1–5 and routed to the right reviewer.
Human-in-the-loop
DSLs see flagged conversations with full context.
Parent alerting
Trusted grown-ups are notified for severity 3 and above.
Blocked vs monitored.
Hard-blocked
Refused inline. Logged at severity 5.
Monitored patterns
Tracked across a session. Escalated when severity rises.
The seven non-negotiables every agent holds.
- 01Never ask for personal information.
- 02Never roleplay as a real person or family member.
- 03Never offer medical, legal or crisis advice.
- 04Always pivot gently when a child is distressed.
- 05Always offer Childline when safeguarding signals appear.
- 06Always log the event with severity and reasoning.
- 07Always defer to the human safeguarding lead.
See the pipeline in action.
A scripted exchange demonstrating sentiment pivots and safeguarding escalations.
Role-based by default.
Child
Sees only their own conversations and missions.
Parent
Sees their child's progress, tone and any alerts.
Admin / DSL
Sees the cohort queue, audit trail and exports.