Real local Qwen inference · Real Bridge Intent · Simulated device.
This 75-second replay is reconstructed from saved execution traces. It is not an Android screen recording; physical-device acceptance remains unverified.
The model receives the same goal in each scenario. Changed controls and a confirmation dialog require a different decision sequence. When saving is broken, the guard stops repeated actions without claiming success.
| Scenario | Result | Model calls | Evidence |
|---|---|---|---|
| Normal task creation | Verified pass | 3 | Run report |
| Changed controls + confirmation | Verified pass | 4 | Run report |
| Broken save | Inconclusive | 5 | Run report |
These are individual demonstration runs, not a statistical benchmark or a measured human-time saving. Qwen repeated Save in the broken scenario; the system did not diagnose the storage failure. Fifteen automated tests passed. The sample APK was built and signature-verified.