Computer Use reports use one of four evidence classes. The class is part of the qualification boundary and is not descriptive copy.
A live provider used the production Runtime, Computer Use tool, owned fixture, and executor path. Provider qualification additionally requires:
complete/end_turn;Missing evidence is invalid or inconclusive. It is never inferred from fixture state alone.
The live provider and production Runtime ran, but the named failure was injected by a wrapper rather than observed from the real host boundary. intervention-recovery currently belongs here because the wrapper injects user_intervened before the backend and HID-age guard run.
Fault-injection reports are useful regression evidence but cannot satisfy a real-runtime provider qualification cell.
A local protocol server verified provider URL, authentication, model ID, streaming tool calls, tool-result reinjection, error flags, and final semantic state. No live provider credential or network model execution is claimed.
Source, schema, or deterministic harness checks only. The superseded direct real-machine qualification runner was removed. The five-round process-restart runner remains as a non-qualifying soak after its qualification checks moved into the canonical Runtime-backed harness.
The canonical operator commands are:
npm run computer-use -- real-ax npm run computer-use -- real-ax --scenario restart-recovery npm run computer-use -- real-model