Hybrid Multi-Agent Moderation Pipeline

A dual-agent architecture combining real-time binary evaluation with periodic pattern analysis—identifying harm without retaining private message transcripts.

Design principles

• Harm detection without surveillance records
• Real-time evaluation before delivery
• Email escalation with severity—not transcripts
• Uniform coverage across all interaction types

Layer 1 — Binary evaluation

Each message receives a yes/no harm-pattern determination before delivery. No human review, verbatim storage, or plaintext logging.

Layer 2 — Pattern analysis

Every 30 minutes, interaction patterns are analyzed for complex toxicity, coercive dynamics, and sustained harassment.

1. Retrieve recent interaction windows
2. Analyze patterns, not raw text
3. Escalate to guardians via email

Guardian notification

All oversight via email—no application sign-in required. High-severity events trigger immediate notification.

Privacy Policy