Voice evaluation callers must wait for the authorized reply, not merely an acknowledgment

Voice evaluation callers must wait for the authorized reply, not merely an acknowledgment.

A test caller can create apparent bot failures when it advances on the first quiet gap after its own utterance. Frontend acknowledgment audio may precede a backend tool result and its authoritative speech instruction; accepting that earlier audio as the reply causes the next test utterance to interrupt the consequential answer. Likewise, immediately closing after a terminal tool result can truncate the final explanation.

For adaptive audio tests, preserve the ordering of caller sends, tool results, speech directives, received media and close events. Require post-directive audio, use a conservative duration/quiet guard where the platform has no playback-complete event, and let the bot's bounded terminal close finish before falling back to client close. Record estimated timing as estimated. On an unexpected state, repair through an explicitly bounded branch or halt dependent steps as unassessed. Keep the original failed run; a corrected test is a new observation.

Separate business invariants, routing progress, delivered audio and conversational quality. A correct persisted review bid does not establish that the caller heard its review-only qualification. Passing backend checks cannot erase missing audio evidence. Conversely, an interrupted reply in a flawed driver does not isolate a production speech defect.