Do not use a context ACK's end as a speech cutoff

When GPT-Live context acknowledgements estimate a delivery interval, do not treat the interval's end as the first possible spoken word. A canonical first word can share that whole interval; filtering it out makes the next correct word appear inconsistent. For a transcript monitor, retain overlapping fragments from the acknowledged interval's start and apply the unchanged canonical lexical check. Reject strictly earlier fragments, validate both interval edges, and preserve current directive and epoch authorization. This changes text classification, not PCM timing or playback proof. Independent recording checks remain necessary, including for audio discarded before acknowledgement.

Applies when: A GPT-Live speech monitor rejects correct replies at their first words. Keywords: OpenAI GPT-Live acknowledgement transcript boundary interval speech monitor PCM