Guardrails

Anti-hallucination, approval, and review controls

The system blocks unsafe outgoing replies, keeps high-risk actions behind approval, and records reviewable traces for every weak-confidence path.

Pre-send Validator

  • Forbidden persona phrases
  • Unapproved prices
  • Overlong replies

Action Approval

  • Payment confirmation
  • Lifecycle updates
  • High-value enrollments

Review Queue

  • Low confidence
  • Blocked reply
  • Admin audit
  • KB correction

Usage Metering

  • Messages
  • AI tokens
  • Media classification
  • Storage
  • Handoff count

Current Safe Default

Payment confirmation, lifecycle writes, and high-value enrollment changes stay admin-approved until live data proves they are safe.

no silent high-risk writes