Safety

Guardrails · what stops the AI from going off the rails

Eight constraints, each enforced in code (not policy doc), monitored by an eval pillar, and visible in an audit row when it fires.

🛡️Enforced controls

0/8

all live in code

🔏Severity · max

RED

registry + policy = RED-tier

📊Watched by

7 pillars

every guardrail eval-tracked

🔄Rollback trigger

< 95%

auto-revert if pass-rate drops

At a glance

Layered defence · six layers stacked

Layer	Mechanism	Answers