Guardrail: pre-action safety: allow / review / block Calibrated, fail-closed AI-safety service for agents — honest 0-1 confidence, abstains rather than guess.