Determines where an autonomous workflow must pause for human authorization. Use for irreversible, externally visible, financially/legal consequential, security-sensitive, high-blast-radius, or materially uncertain actions. Not for deciding whether tool data itself is trustworthy.
