Week 2 · Lesson 06

Policy gate

Safety core: capability checks before any world mutation. See also Safety & compliance.

Time · 1 day Critical
Learn
  • Actor + action + resource authorization
  • Allowlists for irreversible actions
  • Deny must be logged for audit
Build
  1. policy.can(actor, action, resource)
  2. Examples: agent may draft / media.search; only human (or approve token) may instagram.publish
  3. World-mutating tools call policy first; deny → needs_human event
Checkpoint

Even if the model outputs “publish now”, the tool refuses without approval — and a deny event is stored.

Four notes
  • What problem does this layer solve?
  • What is allowed to call it?
  • What must it never do?
  • How do I observe and test it?