Voice vs Text Testing

A test run can exercise the agent over a real voice call or over a text replay of a conversation, and the two are not competing options: they are different lanes with different jobs. Voice is the high-fidelity lane. It catches everything that only exists in speech (latency, interruption handling, what the agent actually says out loud, telephony behavior), and it is slow and expensive, which makes it right for go-live readiness and quality sampling and wrong for a save gate. Text replay is the fast lane: an LLM replays a stored conversation against the candidate agent with mocked tools, in seconds, cheaply enough to run on every change. Hamming's own native pre-save check tried to be both and was unreliable in each direction (text-based and flaky, voice slow and expensive), which is exactly why the proposal keeps Hamming for voice fidelity and puts the pre-save regression lane in-house.