Safety classifications
One profile. No escape hatch — except a human.
- read (navigation, assertions): always allowed
- write-low (form input, drafts): allowed
- write-high (delete, purchase, transfer): blocked by default — a human can approve one action, live
The runtime enforces a single policy profile (controlled-web).
There is no CLI flag, environment variable, or API argument to weaken it — an
agent can never trigger, skip, or pre-approve a blocked action. A person watching
the cockpit can approve one specific action live, right when it happens; nothing
else can. Sensitive input (passwords, credit cards) is blocked unless explicitly
allowed per-call.