Before you press play
What you’ll find in the video
- An allowed output list prevents invented labels but can still force a wrong selection when a suitable fallback is missing.
- The playground tests include negation, adversarial text, exact-value selection, and agent-trace auditing; they test individual cases rather than establishing broad injection resistance.
- The browser example separates action selection from the generative model used when a task needs text.
Worth knowing
Auto-generated English captions reviewed with Gemini. Results are based on eight synthetic playground tests; evaluation times (92-214 ms) exclude network latency, and zero-hallucination claims do not guarantee decision accuracy.