The log may mean only that TTS produced audio or our server sent it. On a bidirectional phone stream, outgoing media can be queued for playback. A caller interrupts, the application clears buffered media, and the sentence that contained the amount or next step never plays. If the conversation transcript still records the whole generated sentence as something the caller heard, the next model turn can assume a confirmation that never happened.

Twilio's Media Streams protocol is a concrete example. Outbound media is buffered. A server can send a named mark and receive a matching mark after playback. It can also send clear, which empties the buffer and returns marks for media that will not play. So a returned mark by itself needs interpretation against the clear and stream state. A TTS-completed event is even farther from proof of playback. The provider's playback acknowledgment still cannot prove a human listened or understood, but it establishes a better transport boundary.

I would track four states for each response segment: generated text, audio sent, playback acknowledged, and cleared or interrupted. Keep the exact text range associated with each segment and its stream ID. On barge-in, stop sending old audio, clear the queue, reconcile outstanding marks, and rebuild the model's conversational state from the portion plausibly played. Do not append the whole planned response to the audible history. If the cut falls inside a word or a critical amount, treat it as uncertain and repeat the important fact plainly before acting on any supposed acknowledgment.

There is a useful interview follow-up here. If the agent says “your refund is complete” and gets interrupted, should it issue the refund again to make the message true? No. The payment's durable action state and the caller's hearing state are separate. Query the action state, then repair the communication. The screen shows $1.50. Why did the voice agent say $150? deals with a wrong amount in the audio text before TTS. This one can have perfectly correct words and still fail at delivery.