Production audio replay needs transcript-assisted media-clock validation before scoring

General lesson: event-to-transcript joins are not enough to trust production audio clips. For replay benchmarks, first map accepted STT turns back through final-received events to caller VAD spans, then calibrate a per-call event-to-media offset on the authenticated caller channel. Use the stored transcript only as an alignment check or bounded nearby-offset selector; never serialize it into the tested audio model. Keep the full audio-to-decision lane separate from transcript-based attribution, reject calls whose spans exceed the recording, and report the residual low-similarity lane rather than silently dropping it.

Applicability and limits: this method applies when recordings have isolated speaker channels and stored turn transcripts but lack a trustworthy sample-zero timecode. Transcript-assisted selection improves clip identity but does not create human routing gold, and difficult accents/noise can remain low-similarity even when the clip is correct.

Applies when: building an audio-to-action benchmark from production recordings, lifecycle events, and stored STT turns. Keywords: production audio replay media clock caller channel STT alignment benchmark