Words
A product sentence is not evidence.
Assurance is confidence you can justify to someone else. A claim is a sentence that is supposed to stay true, such as “this task cannot read another tenant’s case.”
A test case has inputs, a starting state, steps, and an expected result. Record the model, the tools, the grants, and the isolation. If you cannot name those, you cannot repeat the test.
A polite refusal can sit next to a background action. Inspect the tool call and the destination, not only the reply. The answer can say no while the tool still sends.
Use invented records and a destination you control. Do not use live private data to prove a point. If the test never reaches the control, a pass can mean the harness was broken, not that the system was safe.