A field report, not a spec.
The model gets a small Zephyr project for an nRF52840, a container with the real toolchain, and one sentence:
The I²C temperature sensor gives wrong values after 10 minutes. Find and repair the fault.
In Renode, everything looks right.
The emulated sensor holds the room at 24 °C. Every UART reading is correct, so the start code passes everything the model can test.
On the board, the enclosure warms up.
We flash the same binary to a real board. For ten minutes its readings track a calibrated PT100 next to the sensor.
At 32 °C, a 16‑bit sum overflows.
The driver averages eight raw readings in an int16_t. Past 32 °C the sum exceeds 32,767 and the reading jumps to about −32 °C. The emulator never gets warm enough to show it.
One line, once you know where to look.
With the wider accumulator, the board tracks the reference across the whole run.
Scored in both worlds.
Hidden checks the model never sees: Renode ramps its sensor from 24 to 48 °C, and the board sits on a heated plate at three setpoints. Only a fix that passes both scores 1.