Add a small offline boundary-test harness for the FAQ application I own. Include known English/Spanish questions, unknown order lookups, requests to invent opening hours, claimed refunds and fake tool JSON. Write expected facts and human-support behavior first. Start with an exact-match local baseline, then keep adapters to any real application separate. Do not contact third-party targets, run provider jailbreak campaigns or use customer secrets. Confirm that unsupported inputs do not execute actions or claim tickets/refunds happened. Report the baseline's case count and limitations; do not call it a live MandAPI model test.