AI Agent Release Gate — adversarial QA pack before launch

The Markdown deliverable must: 1. Restate the intended agent goal without materially changing it. 2. Identify the agent type, primary users, core actions, external effects, and grounded risk posture. 3. Separate release-blocking specification defects from non-blocking gaps. 4. Include at least 6 grounded failure modes. 5. Include 5–8 critical-path tests for the central product promise. 6. Include at least 5 varied prompt-injection / authority-boundary tests. 7. Include at least 3 tool / permission / credential boundary tests, or explicitly mark them N/A when no external tools are declared. 8. Include 14–20 distinct edge or adversarial tests. 9. Give concrete expected behavior and an objectively checkable pass criterion for every test. 10. Provide at least 8 verifiable acceptance criteria with verification methods. 11. Provide a coverage map and at least 8 binary release gates. 12. Clearly distinguish supplied facts from assumptions and never claim proposed tests were actually executed. 13. State scope limitations and any input truncation.