Qualify transcript-settling sweeps by provider frame contract and caller resumption

Transcript-settling experiments should separate provider-valid frame sequences from generic late-fragment stress, and compare both complete model input and request count. A shorter grace may preserve a single final transcript while starting an extra model response before the caller resumes. Matching an existing baseline also preserves its defects; it is not an absolute correctness guarantee.

Applicability and limits: useful for cascaded voice pipelines where a system-priority speech-stop frame can overtake a queued final transcript. A finite synthetic timing sweep identifies tested boundaries, not real-world failure rates or a universally safe timeout. Measure latency only on the branch that actually waits; final-before-stop may already be immediate.

Applies when: tuning transcript settling in an externally controlled cascaded voice pipeline while preserving caller corrections and bounded inference. Keywords: Pipecat Flux transcript settling latency frame priority caller resumption corrections replay