Output Safety Gate

Checks the assistant's reply before the customer sees it, a pre-output policy layer. Jev asks three different questions in one call, each with criteria: does the reply harm (unsafe tips, dosing, disabling safety gear), leak data the user isn't entitled to (another customer's address, internal notes, keys), or break the company policy (which is in the state)? Code blocks the reply if any flag reaches the threshold. The replies are fixed, so only the gate is measured. Against it: no gate, and an LLM judge given the same three rules. Measured: accuracy, bad replies released, good replies blocked, and the threshold sweep on the highest flag.