Post-call voice diagnosis needs boundary evidence

Post-call voice diagnosis needs boundary evidence and completeness markers

Classify streaming-audio failures from compact boundary evidence, not recordings or logs alone. In the live Pipecat/LiveKit path, increment counters, queue high-watermarks, discarded-tail bytes, and capture outcomes without awaiting storage; flush immutable per-utterance and per-call summaries asynchronously. Mark which checkpoints were expected and observed so missing telemetry becomes “analysis incomplete,” never “healthy.” Application-side acceptance can establish submission, not handset playback; require SIP/RTP or downstream recording evidence for later delivery claims. This preserves media-path latency while making local drops deterministic.

Applies when: asynchronous post-call analysis must distinguish local loss, backlog, and downstream uncertainty. Keywords: Pipecat LiveKit audio telemetry post-call queue capture RTP completeness