Device Twin
The AI Device Twin is Aucert's approach to bridging the gap between emulator testing and real-device behavior. It overlays predictive intelligence on emulator execution to identify issues that would only surface on physical devices.
The problem
Emulators are fast and cheap to run, but they differ from real devices in ways that matter for quality:
| Divergence area | Emulator behavior | Real-device behavior | Impact on testing |
|---|---|---|---|
| Performance | Fast CPU, no thermal throttling | Varies by chipset, throttles under load | Timing-sensitive animations and transitions may pass on emulator but stutter on device |
| Touch input | Pixel-perfect mouse clicks | Imprecise finger taps, fat-finger errors | Touch target size issues are invisible on emulator |
| GPU rendering | Software or host GPU passthrough | Device-specific GPU with driver quirks | Shader rendering, transparency, and gradient differences |
| Network | Direct host network, low latency | Mobile network, variable latency, packet loss | Timeout handling and loading states may not be exercised |
| Sensors | Simulated (mock data) | Real hardware with noise and drift | GPS accuracy, accelerometer-dependent features, biometric auth |
| OS variants | Stock Android | Manufacturer skins (Samsung One UI, Xiaomi MIUI) | Custom system dialogs, permission flows, notification behavior |
How the Device Twin works
The Device Twin operates as a sidecar to the Execution layer (L2), processing emulator results through three stages:
1. Calibration
The calibration model learns emulator-to-device divergence patterns from paired test runs — the same test executed on both an emulator and a physical device. By comparing the results, the model learns:
- Which UI elements render differently on specific device families
- Performance timing adjustments (e.g., "this animation takes 1.3x longer on Pixel 7 than on emulator")
- Touch target issues that only manifest with finger-sized inputs
2. Prediction
Once calibrated, the prediction engine processes emulator-only test results and flags potential device-specific issues:
- Adjusts timing expectations for animations and transitions
- Predicts touch target accessibility based on device screen density
- Flags GPU-dependent rendering that may differ on specific chipsets
3. Confidence adjustment
The Device Twin modifies confidence scores passed to the Analysis layer (L3). If the prediction engine identifies a high-risk divergence, the confidence score is reduced, triggering deeper verification through the Verification Cascade.
Current status
MVP scope: The Device Twin is designed but not built in Phase 1. The current pipeline uses direct emulator execution (Android only) without prediction overlay. Device Twin capabilities are planned for Phase 2.
What's built today:
- Direct Android emulator execution via ADB
- Screenshot capture at each test step
- Execution trace recording with timing data
What's planned for Phase 2:
- Paired test run collection (emulator + device)
- Calibration model training
- Prediction engine integration with L2
- Per-device-family confidence adjustments
What's next
- Architecture overview — Full pipeline architecture
- Knowledge Graph — How app context powers testing