Stage voice-stack demonstrations by the claim each experiment can support

A voice-stack demo should separate three claims: fixed-input technical superiority, interactive conversational behavior, and production business impact. Paired audio replay can compare transcription, turn decisions and resulting structured actions; it cannot establish how callers react to different replies. A bounded live roleplay can expose interruption and repair behavior, but cannot estimate abandonment or production success. A later representative pilot is needed for those operational claims. Add a component-swap comparison only when claiming that integration adds value beyond the improved component; it need not block an initial workflow demonstration.

Applicability and limits: useful for choosing the smallest credible proof when several speech components and task policies change together. This is an experiment-design proposal, not evidence that any stack wins.

Evidence/example — 2026-09-21: constructed freight scenario, no customer data: a caller says a dock number, pauses, then corrects it. A replay can establish whether the stored dock and endpoint timing are correct on that input. It cannot establish whether a real caller would repeat, correct, or hang up after a different agent response. Score latency at genuine completed-turn boundaries, and include discarded speculative synthesis and repair attempts in cost per correct completed update. The inspected paired-ASR-lane work supports isolating component effects; the inspected request-fidelity work supports holding serving inputs constant.