The provider returned an error, timed out, or refused the charge — and the agent still told the customer it was done.
The provider fails before writing anything (a 500), stalls past the timeout, or refuses outright (for example a disputed charge).
Language models are trained to be helpful and conclusive; without the provider’s answer in front of them they tend to report success.
The customer is told the refund is on its way. Nothing arrives.
The sandbox injects the error, the timeout and the refusal on the write. The verdict is decided by the final provider state — the reply cannot make it pass.
An advisory semantic check also flags replies that claim a refund or cancellation that the state does not show. It never turns a failure into a pass.
Guard reports the real outcome of every operation — executed, failed, refused — with the provider’s own reason, so your agent can tell the customer the truth.
Verification re-reads the provider after the call; an outcome that can’t be confirmed is marked unverifiable and raises an incident instead of a silent success.